MINDS

MINDS Champions

Meet the leaders building Meaningful, Intelligent, Novel, and Deployable Solutions to transform businesses and society. 

MINDS are pioneers of innovation technology transformations

These organizations are leading the way in applying frontier technologies to reshape business models, unlock new value, and deliver tangible impact across industries.

49

Selected Transformations

30+

Countries

16

Sectors

6

Impact Areas

Case Studies

Learn more about each project from the MINDS Champions.

Cohort: 2 Location: Beijing, China Application Title: AI monitoring and control for renewable energy infrastructure and robotic intervention safety Team Members: Bin Huang, Enmin Wang, Xin Ren, Ge Ge, Angang Wei, Jiazheng Yu The challenge Large-scale renewable energy infrastructure is difficult to operate and maintain. Renewable energy facilities like photovoltaic plants occupy vast areas in remote locations, making manual inspections inefficient. Equipment such as substations, transformers, and switchgear in offshore wind farms must function reliably in harsh environments, often far from shore and exposed to variable weather conditions. Traditionally, inspection and maintenance rely on scheduled manual checks supported by basic sensor data. Infrared imaging may be used to detect overheating, while visible-light cameras help identify surface issues. These methods are effective in isolation, but each captures only part of the picture. Operational data from equipment is often stored separately from inspection imagery, making it difficult to understand how visual anomalies relate to real-time performance. In addition, many facilities deploy inspection robots and automated devices without centralized oversight, resulting in fragmented data and underused assets. As offshore renewable capacity expands, these limitations increase operational risk and slow response times, particularly in situations where human access is restricted for safety reasons. The solution A unified platform for diagnostics, data management and autonomous response China Huaneng Clean Energy Research Institute, with Huaneng Jilin Power Generation and China Huaneng Group, developed an AI-based platform to manage monitoring, fault detection and unmanned equipment across renewable energy sites. The platform brings together inspection data, operational signals and equipment management within a single system. It supports operations teams by providing a consolidated view of asset condition across geographically distributed sites. Rather than treating diagnostics and data processing as separate functions, the system links them through a layered architecture that connects sensors, AI models and operational workflows. This allows issues to be addressed more consistently. How it works From inspection to intervention The platform supports three connected use cases. The first focuses on fault diagnosis for offshore substations. With numerous densely packed electrical devices within the substation, the integration of automated equipment, devices, and algorithms enables automated inspections to cover the primary content of weekly maintenance personnel inspections at offshore booster stations. Huaneng’s system processes images captured by robotic inspection through a deep learning pipeline, enabling it to flag anomalies and pinpoint equipment-level fault locations. Technicians can obtain a more comprehensive view of the system status without needing to visit the substation. The second use case integrates inspection imagery with operational data. Traditional inspections rely on infrared imaging or visible-light cameras, both of which have inherent limitations. The Huaneng system integrates infrared and visible light imaging into a unified diagnostic layer, simultaneously merging visual observations with operational data. When abnormal heat patterns align with irregular load behaviour, the system flags the issue as higher priority. This linkage allows maintenance teams to move from isolated observations to component-level fault analysis, even when data quality varies across sites. The third use case addresses management of unmanned equipment. Inspection robots, sensors and automated devices often operate independently and generate data in different formats. The platform introduces centralized oversight across group, regional and site levels. Advanced data cleansing algorithms are used to remove noise and correct inconsistencies before data is transformed and integrated into the broader system. This supports coordinated scheduling, lifecycle management and more reliable use of automated assets. Across all use cases, edge–cloud collaboration allows data to be processed locally where needed while supporting centralized analysis and control. How it differs from conventional approaches Reducing fragmentation across sites and systems Conventional renewable infrastructure management treats inspection, operations and equipment control as separate activities. Data is collected in silos and manual coordination is required to link findings across teams. Huaneng’s approach reduces this fragmentation by embedding AI into the core operational workflow. This makes it possible to move from detection to response more quickly, particularly in environments where human access is limited. The system also supports autonomous intervention. In emergency scenarios, robotic systems can perform actions such as operating switches or valves in areas that are unsafe for personnel. Impact Data cleansing accuracy increased to 92%, improving reliability of operational datasets Data processing speed increased by 25% through edge–cloud coordination Detection accuracy for small and medium-sized pointer and dial defects exceeded 90% Dual-spectrum diagnostics achieved high precision in identifying equipment anomalies Autonomous robotic intervention enabled safe emergency operations in inaccessible environments  Why it matters Supporting reliable operation of renewable assets As renewable energy infrastructure becomes more distributed and complex, maintaining reliability requires closer integration between data, diagnostics and control. China Huaneng Clean Energy Research Institute’s platform shows how AI can support this integration by linking inspection, operational analysis and unmanned equipment management within a single system. This approach helps reduce response times, improve asset visibility and support safer operations as renewable capacity continues to expand. 
People's Republic of China | Energy
AI monitoring and control for renewable energy infrastructure and robotic intervention safety
China Huaneng Clean Energy Research Institute, China Huaneng Group, Huaneng Jilin Power Generation Co., Ltd
Cohort: 1 Location: Munich, Germany Application Title: AI-based closed-loop HVAC automation for building performance and energy efficiency Team Members: Rahul Chillar (executive sponsor), Alexander Gloning, Carlos-Eduardo Pazos-Zarain, Monica Arnaudo, Paul BaumannThe challenge Managing large commercial buildings is complex. Heating, ventilation and air conditioning systems must balance energy use with occupant comfort across multiple floors and rooms. Conditions change constantly as people move through buildings and weather shifts throughout the day. Traditional building management systems rely on fixed schedules and manual adjustments. Facility managers monitor dashboards and respond when problems arise, often after occupants report discomfort or energy use exceeds expectations. These systems usually operate in silos and lack the ability to adapt continuously to real-time conditions. Maintaining consistent thermal comfort is particularly difficult. Fixed HVAC settings cannot respond quickly to changes in occupancy or external temperature, leading to uneven indoor conditions and unnecessary energy consumption. As buildings become larger and more dynamic, these limitations become more pronounced. The solution An AI-based closed-loop HVAC automation built on existing infrastructure Siemens developed an AI-based closed-loop HVAC automation that optimises building performance in real time without requiring new hardware. The system connects to existing HVAC equipment and sensors, using live operational data to guide control decisions. Rather than replacing building management systems, it sits alongside them and provides an additional layer of intelligence focused on comfort and efficiency. The objective is to support facility managers by automating routine adjustments while maintaining transparency and control. How it works Closed-loop optimisation of indoor conditions The AI system analyses incoming data from building systems and combines it with short-term forecasts. It learns how different areas of a building respond to heating and cooling actions and adjusts settings accordingly. Decisions are made continuously rather than on fixed schedules. HVAC setpoints are updated as occupancy changes or external conditions shift, helping maintain comfort while avoiding unnecessary energy use. The system operates as a closed loop, monitoring outcomes and refining its behaviour over time. Because it runs on an edge–cloud architecture, the system can respond quickly at the building level while benefiting from centralised analytics. How it differs from conventional approaches Moving beyond static HVAC control Conventional building management systems apply predefined rules that change infrequently and depend on manual intervention. Adjustments are reactive and often delayed. Siemens’ approach uses AI to anticipate changes and adjust HVAC operation proactively. By continuously learning from building utilization, the system reduces the need for manual tuning and supports more consistent comfort across spaces. Importantly, this is achieved without replacing existing infrastructure, which lowers the barrier to deployment. Impact 28% increase in comfort compliance in pilot deployments 6.4% reduction in monthly energy costs at pilot sites Why it matters Improving comfort while reducing energy use As buildings account for a large share of energy consumption, improving operational efficiency without disrupting occupants is an important challenge. Siemens’ control centre shows how AI can support this goal by optimising HVAC operation in real time. By working within existing systems and adapting continuously, the platform helps facilities maintain comfort standards while lowering energy costs, supporting more sustainable building operations at scale. 
Switzerland | Energy
AI-based closed-loop HVAC automation for building
Siemens
Cohort: 2 Location: Hangzhou, China Application Title: Agentic AI for materials discovery Application Subtitle: Team Members: Shuming Zhang, Yuping Li, Luyang Zhang, Chenru Duan,Haojun Jia The challenge Scientific discovery in chemistry is often slowed by fragmentation. Experimental chemists, computational modelers and machine learning specialists typically work with different tools and methods. As a result, collaboration can be slow and progress uneven. Many areas of chemistry still rely on quantum mechanical simulations to understand how reactions occur. One commonly used approach is density functional theory (DFT), which models the behaviour of electrons to predict reaction outcomes. While DFT is highly accurate, it is also computationally intensive. Simulating a single reaction pathway can take hours or days. Exploring large reaction spaces or optimising reaction conditions can take much longer. Because of these constraints, innovation in areas such as materials science and pharmaceutical development often depends on manual trial and error. This limits the speed at which new compounds or processes can be identified. The solution A coordinated AI system for chemical research Deep Principle developed a research platform designed to connect computational chemistry, machine learning and experimental design within a single workflow. At the core of the platform is a multi-agent system that mirrors how research teams operate. Different AI agents are responsible for specific tasks. Some handle computational chemistry. Others focus on machine learning, literature analysis or experiment planning. These agents coordinate their work, allowing simulations, data analysis and knowledge retrieval to proceed in parallel rather than sequentially. The system can plan research steps, run simulations, train predictive models and combine results into actionable outputs. Researchers interact with the platform throughout the process rather than receiving results only at the end. How it works Predicting reaction mechanisms directly A key component of the platform is a reaction mechanism prediction model. Chemical reactions proceed through short-lived configurations known as transition states. Accurately identifying these structures is essential for understanding how reactions unfold, but they are difficult and expensive to compute using traditional methods. Deep Principle’s model generates transition-state structures directly using a custom SE(3)-equivariant graph neural network. This type of model is designed to respect the physical symmetries of molecules, such as rotation and spatial relationships between atoms. An optimal transport framework is used to ensure chemically plausible pathways. This approach allows reaction pathways that previously required hours or days of computation to be generated in seconds. Results are presented in an interpretable way, and researchers can refine outputs or explore alternatives using natural language queries. How it differs from conventional approaches Covering the full research lifecycle Most AI tools in chemistry focus on individual tasks, such as predicting reaction yields or screening candidate molecules. Deep Principle’s platform is designed to support reasoning across the full research lifecycle. The system adapts as new data becomes available and updates its recommendations as experiments progress. Because expert reasoning from different disciplines is embedded into the platform, research teams do not need to assemble rare combinations of specialised skills to move projects forward. This allows research efforts to scale without increasing reliance on highly specialised cross-disciplinary roles. Impact More than 50% of materials simulations automated, compared with a near-zero baseline 500–1,000× faster identification of optimal candidates through virtual screening 30–50% reduction in experimental formulation costs within three to six months 2× improvement in reaction prediction accuracy compared with expert judgement Early pilot projects showing validated R&D cost savings and new revenue from material innovations Approximately 50 users across internal and partner R&D teams, with around 20 API calls per user per day Why it matters Reducing bottlenecks in chemical discovery Chemical discovery increasingly depends on the ability to explore large design spaces efficiently. Deep Principle’s approach shows how AI systems can help reduce bottlenecks created by slow simulation cycles and fragmented workflows. By integrating computation, data analysis and experimental planning within a single platform, the system supports faster iteration while maintaining scientific transparency. This can shorten development timelines in materials and chemical research, where progress has traditionally been constrained by computational and coordination limits. 
People's Republic of China | Chemicals, Materials
Agentic AI for materials discovery
Deep Principle
 Cohort: 1 Location: Ningde, China Application Title: Data-efficient industrial AI systems for manufacturing decision intelligence Team Members: AIMS: Yangbing Lou (executive sponsor), Siqi Zhu, Zhiyong Zhang, Guoyang Yan, Tianhua LuCATL: Jun Ni (executive sponsor), Zhiyang Wu, Lin Ma, Chao Guo, Wenjing Jin The challenge Industrial manufacturing environments generate large volumes of data, but the most critical events are often rare. Equipment failures, quality deviations and process anomalies occur infrequently by design. This makes it difficult to learn from historical examples. Before the use of AI, operations at scale relied on manual inspections, scheduled maintenance and engineer-defined control rules. These approaches depend heavily on experience and static thresholds. They work under stable conditions but struggle when processes fluctuate or when early signals of failure are subtle. In battery manufacturing, where quality consistency is essential, small variations can lead to downstream losses that are hard to detect in time. Conventional machine-learning systems are not well suited to these conditions. They typically require large datasets and frequent retraining, which are rarely available in high-stakes industrial settings.The solution A hybrid AI platform designed for limited data CATL partnered with AIMS to deploy an industrial AI platform designed to operate where data is scarce and reliability is critical. The platform combines physics-informed models with expert process knowledge and machine learning. Physics-based constraints ensure predictions remain consistent with known process behaviour. Expert rules encode operational understanding developed over years of manufacturing practice. Machine-learning components focus on identifying deviations and trends in real-time data. This hybrid design allows the system to generate useful insights without relying on large volumes of labelled failure data. How it works Learning from few examples in real time The system uses few-shot learning to detect rare events. This allows it to recognise early warning signals even when historical examples are limited. Digital twins simulate process behaviour under different conditions, helping predict how changes upstream may affect quality downstream. Decisions are made close to the production line using edge AI. This reduces latency and allows the system to respond quickly to changing conditions. Operators receive alerts and recommendations through existing interfaces rather than new standalone tools. The platform integrates with CATL’s existing manufacturing systems, avoiding disruption to established workflows.How it differs from conventional approaches Moving beyond data-heavy models Traditional AI approaches depend on large datasets and offline analysis. In contrast, the CATL–AIMS solution is designed to operate with incomplete and noisy data. By embedding physical constraints and expert knowledge directly into the model, the system remains interpretable and stable. It adapts continuously without requiring constant retraining. This makes it suitable for production environments where downtime and false alarms carry real cost. Impact 50% reduction in quality variation 50% decrease in operator workload Faster production cycles achieved through earlier detection and process adjustment Why it matters Supporting reliable decision-making in complex manufacturing Manufacturers increasingly need systems that can support decisions when data is limited and consequences are significant. The collaboration between CATL and AIMS shows how hybrid AI can complement human expertise rather than replace it. By focusing on data efficiency and real-time operation, the platform supports more consistent production outcomes without requiring major changes to existing infrastructure. 
People's Republic of China | Energy Technologies
Data-efficient industrial AI systems for manufacturing decision intelligence
AIMS, Contemporary Amperex Technology (CATL)
Cohort: 1 Location: New Taipei City, Taiwan, China Application Title: Self-learning AI factory systems for electronics manufacturing optimization Team Members: The challenge Electronics manufacturing is characterised by high volumes and frequent changes in product mix. Factories must continuously adjust machine settings, production schedules and quality controls as designs evolve and demand fluctuates. While automation is widespread, many operational decisions still depend on the experience of individual technicians and engineers. In processes such as injection molding or printed circuit board assembly, parameter tuning is often performed manually, guided by prior experience rather than systematic feedback. When defects occur, root cause analysis typically happens after the fact, requiring engineers to review logs and test adjustments sequentially. Knowledge accumulates over time, but it is rarely formalised or shared consistently across teams or sites. As production scales, this reliance on experience-driven workflows makes it harder to maintain uniform quality and respond quickly to issues. Conventional improvement efforts focus on lean practices or incremental automation. These methods can raise efficiency, but they do not adapt continuously as operating conditions change, and they struggle to transfer expertise across large, distributed factory networks. The solution A shared AI layer that supports continuous learning Foxconn partnered with Boston Consulting Group (BCG) to develop Project Genesis, an AI platform designed to support self-learning factory operations across electronics manufacturing. Instead of introducing isolated optimisation tools, the initiative establishes a unified AI layer that sits alongside core production systems. This layer captures manufacturing knowledge accumulated over decades and embeds it into AI systems that assist engineers and operators during daily work. The objective is to reduce repetitive manual tuning, shorten problem-resolution cycles and improve consistency, while keeping decision authority with experienced staff. BCG supported the design of the platform architecture and operating model, while Foxconn integrated the system into live production environments across its global operations. How it works Embedding expert knowledge into operational workflows The platform supports several operational modules that share a common data foundation and learning loop. In injection molding, real-time sensor data from machines is analysed to recommend parameter settings that balance quality and throughput. These recommendations are informed by historical outcomes and encoded process knowledge, allowing technicians to make adjustments without relying solely on trial and error. In surface-mount assembly, AI systems monitor production lines and assist with diagnosing defects. When issues arise, the system proposes likely causes based on observed patterns and records corrective actions so that similar problems can be addressed more quickly in the future. For production planning, operators can express scheduling needs in natural language. The system translates these inputs into updated production plans that reflect current constraints and order changes. Across all modules, outcomes are fed back into the platform so that recommendations evolve as conditions shift. How it differs from conventional approaches Reducing reliance on manual optimisation Traditional factory optimisation relies on periodic intervention by experts, with improvements applied locally and often documented informally. As a result, lessons learned in one line or plant do not always transfer to others. Foxconn’s approach embeds AI directly into day-to-day workflows, allowing optimisation to occur continuously rather than episodically. Instead of manually adjusting parameters or diagnosing issues after defects appear, operators receive guidance derived from accumulated knowledge and live data. This reduces variability while preserving human oversight and accountability. Impact 50% reduction in workload related to changeovers 30% decrease in time required to resolve production issues Approximately 10% reduction in overall cycle time Why it matters Supporting consistency in large-scale manufacturing As electronics manufacturing networks expand, maintaining consistent performance across sites becomes increasingly difficult. The collaboration between Foxconn and BCG shows how AI can support this challenge by formalising expert knowledge and applying it in real time during operations. By reducing reliance on manual tuning and accelerating problem resolution, the platform helps factories respond more effectively to variability while allowing experienced staff to focus on higher-value decisions. 
Taiwan, China | Supply Chain & Transportation
Self-learning AI factory systems for electronics manufacturing optimization
Foxconn Technology Group (Hon Hai Precision), Boston Consulting Group (BCG)
Cohort: 2 Location: Beijing, China Application Title: Enterprise retail AI for pricing, loss prevention and energy management Team MembersWumart: Hongyang Wei, Deyu Liu,  Dmall: Nan Song (Nathan), Chaonan Qu (Grace), Xiyu Du (Troy), Pinzhu Chen, Chenyang Zhang (Cheny), Mengdi Li The challenge Modern retail operations depend on thousands of small decisions made continuously across stores. Pricing adjustments, inventory handling, store inspections and energy use all affect margins and customer experience. In many supermarket chains, however, these decisions are still supported by disconnected systems and manual routines. Markdown pricing is often applied uniformly across stores, regardless of local demand or remaining shelf life. Store audits tend to be reactive, relying on staff checks after issues have already occurred. Energy systems frequently run on fixed schedules, even as occupancy and refrigeration needs change throughout the day. These practices can lead to waste, higher operating costs and uneven execution across locations. As store networks expand, coordinating these activities consistently becomes more difficult without real-time visibility and integrated decision support. The solution A unified operating platform for daily retail decisions Wumart partnered with Dmall, a retail digitalisation company, to build a unified operating platform that connects sales data, inventory systems, cameras and in-store sensors into a single workflow. The platform replaces static rules with event-driven processes that adjust throughout the day. Instead of treating pricing and store monitoring as separate functions, the system links them through shared data and coordinated responses. Dmall provides the digital infrastructure and AI capabilities, while Wumart integrates the platform into its store operations and decision-making processes. This allows insights generated centrally to be applied consistently at the store level. How it works Linking data to action in real time For pricing, the system analyses sales velocity, remaining shelf life and current inventory to calculate store-specific discount curves for fresh goods. Prices are adjusted dynamically rather than according to fixed markdown schedules.  For in-store operations, computer vision models monitor video feeds to identify operational issues. These include empty shelves, checkout errors or safety-related incidents. Detected events are converted into guided tasks, which are then verified through follow-up checks or additional sensor data before being closed. For energy management, IoT gateways and virtual meters collect data from refrigeration units, lighting systems and HVAC equipment. The platform forecasts demand and adjusts controls in near real time, balancing energy use with food safety and customer comfort requirements. The system runs on a hybrid architecture, combining cloud coordination with edge execution inside stores. This supports fast response times while maintaining local control over data. How it differs from conventional approaches Moving beyond manual coordination Traditional retail operations rely on predefined rules and periodic reviews. Decisions are often made in isolation by different teams, with limited feedback between pricing, store execution and energy management. Wumart and Dmall’s approach coordinates these activities through a shared operating layer. Decisions that were previously manual or scheduled are triggered by real-time events. This reduces delays and helps ensure more consistent execution across stores. Rather than adding new tools for staff to manage, the platform integrates into existing workflows and translates insights into concrete tasks. Impact Pilot deployments across approximately 100 stores, supporting pricing and operations teams Reduction in deep markdowns, with 20% of stores no longer requiring 50% discounts and improved full-price sell-through 85% reduction in self-checkout shrink, alongside lower labour costs and faster issue resolution Approximately 26% reduction in daily energy use, saving around 1.3 million kWh annually Avoidance of roughly 85,600 tonnes of CO₂ emissions per year across pilot sites  Why it matters Supporting consistent retail execution at scale Retail performance depends on how well daily decisions are coordinated across stores. The collaboration between Wumart and Dmall shows how integrated data and AI-supported workflows can help reduce waste, improve margins and manage energy use more effectively. By connecting pricing, operations and energy management within a single system, the platform supports more consistent execution without increasing operational complexity. This approach becomes increasingly relevant as retail networks grow and operating conditions become more dynamic. 
People's Republic of China | Food & Retail
Enterprise retail AI for pricing, loss prevention and energy management
Dmall, Wumart Stores
Location: Beijing, China Application Title: Policy-aware AI forecasting for electricity, carbon, coal and certificate markets Team Members: Wenqiang (Executive Sponsor), Xiao Liu, Wei Xu, Enjun Wang, Shusen Guo and Jianliang Yang. The challenge Energy markets are among the most complex and fast-moving parts of the global economy. Electricity prices, carbon allowance values, coal indices and renewable energy certificates can shift within hours in response to policy announcements, dispatch instructions or regulatory changes. Unlike many financial markets, energy markets are shaped as much by government decisions as by supply and demand. New emissions rules, price caps, trading mechanisms or grid directives often alter incentives before their effects appear in numerical market data. Traditional forecasting tools rely primarily on historical price and volume data. As a result, they struggle to respond when policy interventions drive sudden changes, leading to inaccurate forecasts, higher financial risk and less efficient power scheduling. For grid operators, generators and trading desks, these gaps can translate into increased compliance costs, missed revenue opportunities and difficulty aligning dispatch decisions with evolving regulatory requirements. The solution Forecasting that integrates policy and market data The National Institute of Clean-and-Low-Carbon Energy (NICE); part of China Energy Investment Corporation Limited, developed an AI-powered forecasting system designed to combine numerical market analysis with policy understanding. The system pairs a large language model specialised in energy policy and market regulation with a multivariate time-series forecasting network. Through a retrieval-augmented architecture, it continuously scans licensed policy databases, regulatory bulletins and trading reports, linking this information to live market data such as electricity prices, carbon allowance trades and coal indices. This allows the model to assess how new rules, policy signals or regulatory adjustments are likely to influence market behaviour before those effects are fully reflected in price movements. How it works Linking text, time and transactions The system operates on on-premises GPU clusters and is used by analysts, traders and planning teams for short- and medium-term forecasting, carbon cost estimation and dispatch planning. At a technical level, the platform combines retrieval-augmented generation (RAG), temporal text vector databases and multivariate transformer decoders. Rather than treating policy documents as static context, each document is timestamped, verified and processed through a retrieve–filter–extract pipeline to ensure relevance and credibility. Training data includes: Approximately 1,000 authoritative policy and technical documents Around 3,000 time-sensitive reports from financial and regulatory sources More than 200,000 rows of transaction data across electricity, coal, carbon and green-certificate markets The model uses LoRA-based fine-tuning to adapt to the energy domain, reducing training energy use by 95% compared with full retraining. How it differs from conventional approaches Dynamic alignment between policy and prices Conventional time-series forecasting models treat text and numerical data as loosely connected inputs, often aligned only during initial training. NICE’s system instead performs dynamic temporal retrieval, continuously sourcing and aligning new policy and market information as it emerges. A bidirectional alignment mechanism allows regulatory insights to inform numerical predictions and, in turn, allows market behaviour to update how policy language is interpreted. This creates ongoing synchronization between language and data. As a result, forecasts are not only more responsive but also explainable. Each prediction can be traced back to the specific policy announcement, regulatory change or market event that influenced it. Impact Price forecasting accuracy improved, with error rates reduced to 0.69% MAPE for weekly electricity forecasts and 2.68% for daily forecasts Carbon price forecasting accuracy of 1.21%, and 1% for coal markets 50% reduction in price forecast error compared with existing methods 95% reduction in GPU energy use through LoRA-based fine-tuning 1% increase in forecast-driven trading optimisationApproximately USD 13.7 million in additional annual trading revenue 3–5% reduction in carbon cost exposure through improved compliance alignment Deployed across 300 internal users in four provinces Why it matters Improving decision-making in policy-driven markets Energy markets increasingly operate at the intersection of economics and regulation. NICE’s approach shows how AI systems that understand both policy language and numerical data can support more accurate forecasting and more transparent decision-making. By linking forecasts directly to the policy and market signals that drive them, the system helps analysts and planners respond earlier to change, reduce exposure to regulatory risk and align trading and dispatch decisions with evolving low-carbon objectives. 
People's Republic of China | Energy
Policy-aware AI forecasting for electricity, carbon, coal and certificate markets
National Institute of Clean-And-Low-Energy (NICE), China Energy Investment Corporation
Cohort: 1 Location: Zug, Switzerland Application Title: Industrial AI framework for quality control and operational excellenceTeam MembersEthonAI: Julian Senoner (Executive Sponsor), Bernhard Kratzwald, Davide Vecchione, Hendrik Schmid, Arman Pour Tak Dost.Siemens: Peter Koerte (Executive Sponsor), Sébastien Bey, Alexander Dierolf, Andreas Duerr, Matteo Odermatt, Sascha LeadbetterThe challenge Manufacturing relies on consistent quality to control costs and meet customer expectations. Yet even in highly automated factories, a large share of available production data is never analysed. Sensors generate continuous streams of information about machines, processes and outputs, but much of this data remains unused. Quality control often depends on manual inspection or narrowly scoped automation. Visual checks are set up line by line and tuned locally. When defects appear, engineers typically rely on experience-driven root cause analysis supported by small data samples. These approaches can identify obvious problems, but they struggle to keep pace with the volume and complexity of modern production data. As a result, inefficiencies persist, scrap increases and improvement efforts remain difficult to scale across sites. The solution A standardised AI platform for factory-wide quality control Siemens partnered with EthonAI to introduce a unified industrial AI platform for quality control and operational analysis. The platform is integrated with Siemens Industrial Edge, allowing it to run directly within factory environments. Rather than deploying separate inspection systems for each production line, Siemens uses a shared platform that can be applied consistently across factories in different regions. EthonAI provides the AI models and analytics framework, while Siemens provides domain knowledge and integrates its industrial software solution. This enables rapid rollout without redesigning local systems.How it works From inspection to root cause insight The platform ingests machine data, process signals and quality measurements from existing factory data platforms. AI models analyse this information continuously to detect deviations and emerging defects. For visual inspection, the system can learn from a small number of reference images taken under stable conditions. This allows inspection setups to be completed in minutes rather than days. Once deployed, inspections run automatically and flag anomalies in real time. Beyond detection, the platform applies causal analysis to link quality issues to underlying process conditions. Engineers can move from observing a defect to understanding contributing factors without switching tools. Insights are translated into alerts and recommendations that can be acted on directly. How it differs from conventional approaches Moving beyond isolated inspections and manual analysis Traditional quality programmes rely on periodic audits and expert-led analysis of limited datasets. While effective in controlled settings, these methods cannot continuously interpret large-scale sensor data across multiple plants. The Siemens–EthonAI approach treats quality data as a continuous learning source. By analysing full data streams and applying manufacturing-specific models, the system identifies patterns that manual sampling would miss. Standardising inspection and analysis through a single platform also reduces duplication and supports consistent practices across sites.Impact More than $10 million in projected savings Deployment across 20 factories, with further expansion planned Inspection setup time reduced from days to minutes Decrease in customer complaints due to early defect detectionWhy it matters Scaling quality improvement across factories As manufacturing networks grow more complex, maintaining consistent quality requires tools that can operate at scale. The collaboration between Siemens and EthonAI shows how industrial AI can support this need by combining automated inspection with deeper process insight. By standardising quality analysis across factories, the platform helps organisations address inefficiencies earlier while reducing the effort required to deploy and maintain inspection systems. 
Switzerland | Energy
Industrial AI framework for quality control and operational excellence
Siemens, EthonAI
Cohort: 2 Location: Rockville, Maryland (US) and Bangui, Central African RepublicApplication Title: Private LLM infrastructure and edge intelligence for urban maintenance and safety Team Members: Samuel Alemayehu (Executive Sponsor), Olivier Nizeyimana, Abbai Belai, Florent BodoThe challenge Across rapidly growing cities, particularly in parts of Africa, infrastructure systems face persistent challenges. Roads deteriorate faster than they can be repaired. Construction sites operate under uneven safety oversight. Municipal teams responsible for maintenance and inspection are often small and have limited access to digital tools or specialized engineering expertise. Much of this work falls to mid-skill professionals; inspectors, technicians and site coordinators, who are expected to manage complex assets with limited support. At the same time, many smart-city solutions are designed for high-income markets. They rely on centralized cloud platforms, assume stable connectivity, and place control of data and workflows outside local institutions. In practice, this makes them difficult to localize and harder to govern.The solution Locally deployed AI for infrastructure oversight Cambridge Industries focuses on deploying AI systems that operate within locally controlled environments and are accessible to people working directly in the field. The company has developed a private large language model (LLM) infrastructure stack that runs entirely within client-owned systems. This approach allows municipalities and contractors to retain control over data, workflows and configuration. Complex engineering standards and safety regulations are translated into mobile-friendly tools that inspectors and site managers can use on site, supporting more consistent decision-making without requiring specialized training. Rather than centralizing intelligence in remote platforms, the system is designed to function where infrastructure is built and maintained. How it works AI tools embedded in day-to-day workflows The platform is implemented through two primary applications. Orbit is a predictive road-maintenance tool that enables inspectors to capture road-surface images using smartphones. AI models analyze these images to identify cracks and early-stage deterioration, then generate colour-coded heatmaps that indicate where maintenance is most urgent. This allows municipalities to prioritize interventions before minor defects develop into more extensive repairs. In pilot deployments across Nigeria (Butterfly Island), Bangui (Central African Republic), and Lualaba (DRC); in partnership with large-scale developer partners, approximately 150 inspectors submit around 20 inspection captures per week per municipality. The system processes more than 12,000 AI queries per month, and early pilot results indicate up to a 40% reduction in emergency road-repair costs within six months.The second application, Ndege, supports construction-site safety monitoring using autonomous drones. Drones collect imagery that is analysed to identify missing barriers, protective equipment or signage. Alerts are sent directly to site managers via WhatsApp or dashboard interfaces, allowing issues to be addressed quickly. At pilot sites, the system records around 400 drone flight hours per month across 20 operators, has generated more than 3,000 AI-generated safety alerts, and has contributed to a 50% reduction in recorded incidents based on site logs over the pilot period.How it differs from conventional approaches Keeping intelligence local Both Orbit and Ndege are built on private multimodal AI models that combine computer vision with fine-tuned LLMs derived from LLaMA-family and other Open Source architectures. The models are trained on locally collected data, including smartphone footage, drone imagery, inspection logs and more than 100 standardized engineering and safety documents/manuals/checklists. All models run on client-owned servers with encrypted data flows and role-based access control. This avoids reliance on external cloud platforms and allows systems to be adapted to local regulatory, operational and connectivity conditions. By embedding decision support into smartphones, drones and edge devices, the platform supports decentralized use without removing oversight from local institutions. Impact Pilot results indicate up to 40% reduction in emergency road-repair costs over a six-month period (pilot sites) Annualized, pro-rated maintenance costs reduction from approximately $5 million to $3 million per municipality based on pilot spend patterns, with additional scale efficiencies anticipated as more of the road network is coveredOver 12,000 model inferences processed per month across the piloted road network(s)Approximately 400 drone flight hours per month across pilot sites More than 3,000 AI-generated safety alerts issued during the pilot periodReported up to 50% reduction in recorded safety incidents during the pilot period (based on site logs)Why it matters Supporting more reliable infrastructure maintenance Cambridge Industries’ approach illustrates how locally deployed AI can support infrastructure maintenance and construction safety in contexts where resources and expertise are limited. By keeping data and decision-making close to the field, the system helps municipalities identify issues earlier, prioritize interventions and reduce reliance on emergency repairs. This supports more consistent maintenance outcomes while maintaining local control over critical infrastructure systems. 
USA, Central African Republic | Engineering & Construction
Private LLM infrastructure and edge intelligence for urban maintenance and safety
Cambridge Industries Group
Cohort: 1 Locations: Beijing & Suzhou, China Application Title: AI-driven battery cell design and simulation democratization Team Members: Ke (Luke) Hu (executive sponsor), Zhe Li (executive sponsor), Lance Zhao, Weinan Zhou, Linghao Lu, Qiyu Chen, Zhixuan Wu, Liming Chang The challenge Battery cells are central to electric vehicles and energy storage, yet designing them is a slow and specialised process. A single cell must balance energy density, power output, safety and lifespan, while also meeting manufacturing constraints. Small changes to materials or structure can affect performance in unexpected ways. Traditionally, battery design relies on a combination of physical testing and design-of-experiments methods. Engineers adjust parameters, build prototypes and test them repeatedly. This process can take several years and consumes significant material and energy. Simulation tools have improved efficiency, but they often depend on expert interpretation and are disconnected from real manufacturing constraints. As a result, optimisation remains slow and difficult to scale, particularly for smaller teams or new entrants. With global demand for batteries rising rapidly, these limitations have become a bottleneck for innovation. The solution A digital platform for predictive battery design Tsinghua University partnered with Electroder to develop an AI-driven platform that brings advanced battery design tools into a single digital environment. The platform integrates physical models that describe how battery materials and structures behave with AI-based optimisation methods. This allows users to explore design options virtually before committing to physical prototypes. Instead of relying on trial-and-error testing, engineers can define performance goals and constraints, then evaluate how different design choices affect outcomes. Electroder provides the platform infrastructure, while Tsinghua University contributes domain expertise in battery science and modelling. How it works Simulating cells before they are built The system uses multi-physics models to represent processes at the electrode and cell level. These models capture electrochemical behaviour and material interactions that influence real-world performance. AI algorithms then search through the design space, adjusting variables such as composition and geometry within manufacturable limits. Users can specify objectives such as higher energy density or longer cycle life. The platform runs simulations to assess trade-offs and identify promising designs. Sensitivity analysis highlights which parameters matter most, helping teams focus effort where it has the greatest effect. A built-in assistant allows users to upload scientific papers or internal documents and extract relevant design insights conversationally, lowering the barrier for teams without deep modelling experience. How it differs from conventional approaches Reducing reliance on physical trial cycles Conventional battery development depends on repeated physical prototyping. Even when simulations are used, they often serve as advisory tools rather than drivers of design decisions. The Tsinghua–Electroder platform shifts this balance by allowing most exploration to occur virtually. Physical testing is still required, but it is used to validate a smaller number of well-informed designs. This reduces material consumption and shortens development timelines. By embedding manufacturing constraints into optimisation, the platform also helps ensure that simulated designs can be produced in practice. Impact Concept-to-prototype cycles accelerated by 3.6× 40% reduction in material and energy waste during development Why it matters Broadening access to advanced battery R&D As battery demand grows, innovation needs to move faster while reducing waste. The collaboration between Tsinghua University and Electroder shows how AI-based simulation can help address this challenge by making advanced design capabilities more accessible. By reducing dependence on long trial cycles, the platform supports more efficient development and lowers barriers for teams working on next-generation battery technologies. 
People's Republic of China | Energy Technologies
AI-driven battery cell design and simulation democratization
Electroder, Tsinghua University
Cohort: 2Locations: Gentilly, France & Munich, GermanyApplication Title: Agent-powered AI transformation engine accelerating time from discovery to patient impactApplication Subtitle: A multi-agent transformation approach that embeds advanced analytics intoeveryday decision-making across biopharma R&D and operations, shortening the path from discovery to patient accessTeam MembersOAO: Dr. René Eber, Dr. Julian Schirmer, Eirik Sjåholm Knudsen, Marcel HolzapfelSanofi: Emmanuel Frenehard, Julien Marcaut, Andrew Philips, Fabrice Collet, Joanne BeardsleyThe challengePharmaceutical research and development generates vast amounts of data and scientific insight.However, in large global organisations, the processes that turn discovery into patient impact often move more slowly than the science itself.Expertise is distributed across many functions. Discovery scientists, clinical teams, manufacturingspecialists and commercial staff all work with different tools and decision frameworks. Thisfragmentation makes it difficult to share insights and coordinate action. Traditional digitaltransformation programmes tend to focus on deploying new technologies or delivering isolated usecases. While useful, these efforts often leave everyday ways of working unchanged and limit how widely AI is adopted across the organisation.As a result, opportunities to shorten development timelines or improve operational decisions canremain underused. The solutionOrganisation-wide AI adoption supported by agentsSanofi, working in partnership with OAO, is addressing this gap by focusing on how AI is used across the organisation, rather than where it is deployed.OAO has been deploying its multi-agent system across all hierarchy levels, serving as the enterprise's transformation engine. The AI agents create a bidirectional flow: capturing ground-truth insights from employees, converting their domain expertise into investable innovations, and steering leadership to place focused AI bets. This turns employees from transformation recipients into active participants.The joint initiative aims to make AI part of daily work for employees across functions and seniority levels. Agents help teams identify opportunities where AI could add value, translate employee input into concrete use cases, and connect local experimentation with enterprise-level priorities. This organizational intelligence helps leadership focus investment on the most promising initiatives, including scaling priority use cases in commercial operations and biologics discovery.This makes AI adoption a participatory process rather than a centrally defined rollout. How it worksConnecting AI across research and operationsThe programme spans Sanofi's full value chain, from research & development to manufacturing & supply to commercial operations. The AI agents have generated thousands of use case ideas across the organisation. Several of these have been prioritised by leadership and scaled enterprise-wide, including two flagship initiatives in biologics discovery and commercial operations.In biologics discovery, a unified AI platform combines predictive and generative models with automated data ingestion and retraining. Scientists can build and deploy models in weeks rather than months, reducing the time required to explore hypotheses or analyse protein sequences. Processing that once took days can now be completed much faster, enabling near-real-time predictions for a large share of active projects.In commercial operations, an AI-driven omnichannel system supports engagement with healthcareprofessionals. Deployed across multiple markets and brands, the system uses a set of specialised models to recommend actions tailored to local context. These recommendations are integrated into existing workflows, allowing teams to act on insights without adopting new tools or processes.Learning from scientific and commercial use cases feeds back into the broader system, allowingmethods developed in one area to inform others. In parallel, the AI Agents platform continuously captures barriers and opportunities from employees and leaders and translates them into prioritized initiatives – shaping where these research and commercial systems are scaled and invested in. How it differs from conventional approachesTurning employees into active transformation participants, not recipientsConventional AI programmes in large enterprises often prioritise individual applications or technology platforms. Sanofi and OAO’s approach focuses instead on embedding AI into decision-making and collaboration.  Rather than requiring employees to adapt to new systems, AI agents meet employees where they are – surfacing their domain expertise and turning it into actionable innovations and initiatives. This lowers the barrier for participation and, over time, builds organisational capability and shared understanding of where AI is most useful.The emphasis is on sustained adoption rather than one-off implementation. Impact·        More than 1,300 AI use cases co-created through organisation-wide participation·        Over 1,400 employees actively engaged to unleash AI innovations·        450 data-informed initiatives shaped by management and 10+ executive AI bets initiated by leadership·        20% AI readiness uplift with transformation efforts and improved clarity around AI priorities·        Single-digit commercial performance uplift, with a 10:1 return on investment across deployed use cases·        11 AI models in production supporting biologics discovery, with model build times reduced from several months to under three Why it mattersReducing the gap between discovery and patient accessSanofi and OAO’s collaboration highlights the role organisational design plays in effective AI adoption. By focusing on participation and integration into everyday work, the initiative supports more consistent use of AI across R&D and operations, ultimately for accelerating delivery to patients This approach helps large biopharma organisations like Sanofi move insights into practice more efficiently, supporting faster decision-making without relying on narrow centres of expertise. Over time, this can contribute to further shortened paths from scientific discovery to patient benefit. 
France, Germany | Health
Agent-powered AI transformation engine accelerating time from discovery to patient impact
Sanofi, OAO
Cohort: 1Location: New York, United StatesApplication Title: Automated AI platform for digital transformation and process migrationTeam Members: Rohit Kapoor (executive sponsor), Deepti Kalra, Anand Logani, Shantanu Paliwal, Shubham Jain, Swati Malhotra The challengeLarge organizations rely on software platforms that were built decades ago. Systems based on technologies such as SAS or legacy Oracle environments often remain critical to daily operations, even as organizations move toward cloud infrastructure.Migrating these systems is difficult. Documentation is incomplete. End-to-end data and code lineage is often missing. Dependencies across jobs, tables, reports, and downstream applications are poorly understood. Redundant, unused, or overlapping logic has accumulated over years of incremental change. Specialist knowledge sits with a small number of experienced engineers.As a result, migration projects are slow, expensive, and risky. Much of the effort is spent on manual code assessment, dependency discovery, redundancy analysis, and rewriting, with teams working through long sequences of repetitive tasks. Late-stage defects; often discovered during testing or parallel runs, can lead to significant rework and extended timelines.Traditional automation tools help with isolated steps, but they do not address migration as an integrated, end-to-end process. The solution An AI platform built specifically for platform migration EXL developed an automated AI platform designed to support end-to-end platform migration, from estate assessment through conversion, testing, and stabilization. Rather than relying on a single general-purpose model, the platform uses multiple specialized AI agents, each focused on a specific stage of the migration lifecycle. These agents are trained using EXL’s historical migration data, capturing patterns related to code structure, dependency resolution, data transformations, and validation logic that are rarely documented formally.The goal is not to remove human oversight, but to reduce the manual effort required for assessment, rationalization, conversion, testing, and optimization—while keeping experts involved in key architectural and business decisions. How it works Orchestrated agents with human review Each AI agent performs a defined role. One analyses legacy code and data structures to extract metadata, lineage, and usage patterns. Another maps dependencies across systems, applications, and reports, identifying redundancies and opportunities for simplification. Other agents handle code conversion, test-case generation, reconciliation, and performance tuning.The agents operate within an orchestration layer that sequences tasks, manages hand-offs, and enforces validation checkpoints. For readers unfamiliar with this approach, the system functions much like a coordinated engineering team. AI handles repetitive, time-consuming analysis and transformation work. Human reviewers focus on design choices, exceptions, risk areas, and final approvals.The platform supports a wide range of legacy technologies and target environments. It integrates securely with existing code repositories and data platforms, produces auditable lineage and validation artifacts, and runs on a cloud-agnostic infrastructure. Human-in-the-loop controls ensure outputs are reviewed where business or regulatory risk is highest. How it differs from conventional approaches Reducing reliance on manual rewrites Conventional migration projects depend heavily on engineers manually analyzing dependencies, rewriting code, and validating outputs. Automation, where used, is typically limited to narrow tasks and often lacks consistency, transparency, or test coverage.EXL’s approach treats migration as a coordinated, lifecycle-driven process rather than a sequence of disconnected steps. By fine-tuning models on prior migration work and using task-specific agents, the platform systematically addresses assessment, lineage, redundancy elimination, conversion, testing, and optimization together.This reduces rework, improves consistency, and embeds validation throughout the migration rather than deferring risk to the end. Performance optimization and cost efficiency are treated as integral outcomes, not post-migration afterthoughts.Impact ·        More than 80% reduction in manual effort for code assessment, dependency discovery, and lineage analysis·        70–80% reduction in manual effort for code conversion and validation·        20–40% reduction in overall project cost·        30–40% improvement in post-migration code and query performance Why it mattersMaking large-scale modernization more predictable As organizations modernize critical systems, migration speed and reliability matter as much as technical accuracy. EXL’s platform demonstrates how AI can support this transition by reducing manual workload, improving transparency through lineage and testing, and preserving expert control.By shortening timelines, improving validation, and eliminating hidden dependencies and redundancy, the approach helps organizations modernize legacy platforms with less disruption and more predictable, auditable outcomes.
USA | Energy
Automated AI platform for digital transformation and process migration
EXL Service
Cohort: 2 Location: Suresnes, France Application Title: AI-Designed Phage Therapy for Antibiotic-Free Animal Health Application Subtitle: An AI-guided platform that predicts phage–bacteria interactions from genomic data to rapidly design adaptive phage therapies, replacing antibiotics in livestock farming. Team Members: Adèle James, Alexandros Pantalis The challenge Antimicrobial resistance is no longer a distant or abstract risk. It’s already affecting how food systems, farming and public health intersect. Nearly three-quarters of the world’s antibiotics are used in livestock farming. Over time, that scale of use has done what biology always does: it has adapted. As antibiotics lose their effectiveness, farmers face rising mortality rates, tighter regulations, and a shrinking set of options to protect animal health. One alternative that has regained attention is bacteriophage therapy. Bacteriophages, often referred to as phages, are viruses that infect and kill specific bacteria while leaving other organisms unaffected. They occur naturally and have been studied for decades. However, turning phages into reliable treatments has proven difficult. Traditional development relies on manual screening to identify which phages work against which bacteria, a process that can take up to two years. In environments where bacterial pathogens evolve quickly, that timeline limits practical deployment.  The solution Predictive phage therapy, designed by AI Phagos approaches the problem from a different angle. Instead of testing phages one by one, its AI-powered platform predicts phage–bacteria interactions directly from genomic sequences, then designs multi-strain phage cocktails that can adjust as bacterial populations change. At the core of the system is a combination of generative AI and machine-learning ranking models to evaluate large numbers of potential phage–bacteria pairings.  Rather than producing a single static solution, these tools enable the continual updating of the phage mix used in the medicine, allowing it to remain effective as bacteria evolve.The platform is trained on a large and diverse biological foundation; thousands of genomes and thousands of validated interactions. This scale allows the model to capture patterns in host–pathogen relationships that would be difficult to derive through manual testing alone. How it works Closing the loop between computation and the lab A central feature of the platform is the feedback loop between computational prediction and laboratory testing. Candidate phage cocktails generated in silico are tested in the lab, and the results are fed back into the model to improve subsequent predictions. The system operates on scalable infrastructure capable of processing genomic data across species and geographies. It is supported by a growing phage bank representing roughly 1% of all known phage genomes. As new experiments are conducted, both the dataset and the model continue to improve. This approach shortens discovery cycles and reduces reliance on repeated trial-and-error screening. How it differs from conventional approaches From reactive screening to anticipatory biology Phagos shifts phage therapy development away from a reactive mindset. Instead of responding to outbreaks after they occur as conventional phage development, the platform anticipates bacterial behaviour and adapts treatment strategies accordingly. By combining AI-driven prediction with experimental validation, the platform provides a structured way to explore phage therapies at scale. This makes development more systematic and less dependent on manual expertise alone, particularly in settings where rapid response is required. Impact Phage therapy development reduced from two years to two months Testing cycles shortened from days to minutes Lab workflows compressed from 48 days to a single day per cocktail Model accuracy reaching 95%, with success rates roughly double those of manual methods Clinical outcomes include: Complete elimination of Salmonella Measurable gains in poultry weight Improved farmer's gross marginsAdoption to date: Three billion-$ enterprise clientsFive active pilots across Europe, Asia and the AmericasWhy it matters An alternative to antibiotics in animal health Phagos demonstrates how AI can support the development of biologics that operate differently from conventional antibiotics. By reducing development time and improving the reliability of phage selection, the platform makes it easier to consider phage therapy as a practical option in animal health. In doing so, it contributes to global efforts to address antimicrobial resistance, not by stretching the lifespan of current drugs, but by supporting the development of treatments that operate differently from conventional antibiotics 
France | Food & agriculture, Chemicals
AI-Designed Phage Therapy for Antibiotic-Free Animal Health
Phagos
Cohort: 1 Location: Hangzhou, China Application Title: Multimodal AI platform for clinical decision support and medical consultation services Team Members: Zhang Junjie (Executive Sponsor), Liu Junwei, Zhu Dongwei, Feng Lei, Wei Peng, Wang JianThe challenge Healthcare delivery is often fragmented. Patients interact with multiple providers across different stages of care, while clinicians rely on separate systems for records, diagnostics and follow-up. As a result, information is spread across formats such as clinical notes, imaging results and sensor data, making it difficult to form a complete picture of a patient’s condition. Many “smart health” systems attempt to address this by adding AI tools to existing workflows. These tools may assist with a single task, such as triage or image analysis, but they rarely connect across the full patient journey. Clinicians must reconcile outputs manually, and patients experience care as a series of disconnected interactions. Privacy and compliance requirements further limit the ability to integrate data across institutions. This fragmentation reduces efficiency and makes it difficult to scale AI support beyond isolated use cases. The solution A unified platform spanning the patient journey Ant Group developed a multimodal AI platform designed to support clinical decision-making and patient consultation across different stages of care. At its core is a medical-grade large language model capable of processing text, medical images, audio and video. This allows the system to reason across symptoms, laboratory results and imaging data within a single framework. A privacy-preserving data layer encrypts and harmonises clinical information from hospitals, primary care clinics and consumer health devices, enabling secure use in regulated environments. The platform assists both clinicians and patients. Before a visit, an AI assistant helps patients record symptoms and schedule appointments. During the visit, it provides convenient medical treatment for users throughout the entire process, and offers customized intelligent triage services to enhance the user's healthcare experience. Following diagnosis and treatment, the system also delivers functions such as follow-up care, medication guidance, and rehabilitation planning services, thereby establishing a continuous feedback loop rather than a one-time interaction.How it works Edge–cloud coordination with medical-grade governance The platform uses a hybrid edge-to-cloud architecture. Real-time interactions, such as voice-based consultation or digital-human interfaces, run at the edge to minimise latency. Model training and updates occur in the cloud, allowing improvements to be deployed quickly across institutions. The system is built in layers. General foundation models provide base capabilities. A domain-specific layer is trained on professional medical datasets and clinical literature. A top-tier medical model is then refined through reinforcement learning and cross-modal alignment to support clinical reasoning. Governance is handled through a multi-layer security framework that includes standardised evaluation processes and ethical oversight. This design supports use in sensitive contexts while maintaining transparency and compliance. How it differs from conventional approaches Moving beyond isolated clinical tools Traditional health AI systems are typically deployed as standalone applications. Each tool addresses a narrow task and relies on siloed data. Integration is left to hospitals, increasing operational burden. Ant Group’s approach provides a shared service layer that supports multiple clinical functions within a single system. Rather than bolting AI onto existing tools, the platform connects data, reasoning and interaction across the care pathway. This reduces manual reconciliation and supports more consistent use across departments. The hybrid architecture also allows on-premise deployment where required, addressing compliance constraints that limit adoption of cloud-only solutions. Impact Auxiliary diagnostic accuracy above 90% across supported use cases 80% reduction in medical literature search time for clinicians Serving 160 million users, with nearly 1 million doctors across more than 5,000 institutions Why it matters Supporting scalable and trusted clinical AI As healthcare systems seek to improve access and efficiency, AI tools need to integrate into care delivery rather than operate at its edges. Ant Group’s platform shows how multimodal models, combined with privacy-first infrastructure, can support clinical reasoning across the patient journey. By reducing fragmentation and supporting continuous interaction between patients and clinicians, the approach offers a pathway toward more scalable and consistent AI-supported care. 
People's Republic of China | Health
Multimodal AI platform for clinical decision support and medical consultation services
Ant Group Co., Ltd
Cohort: 2 Locations: Santa Clara & Sunnyvale, United States Application Title: Reinforcement learning, generative AI and agentic AI for chip designApplication Subtitle: A design framework that applies reinforcement learning and agent-based systems within existing EDA tools to coordinate simulations, reduce manual execution work, and support complex chip design under tight time and talent constraints. Team MembersSynopsys: Shankar Krishnamoorthy (executive sponsor), Thomas Andersen, Arun Venkatachar, Anand ThiruvengadamAMD: Mydung Pham, Rajit Seahra.  The challenge Designing modern semiconductor chips has become increasingly complex. Chips must deliver higher performance and better energy efficiency, even as design timelines shorten and engineering teams remain constrained. At the same time, the global shortage of semiconductor engineers is expected to widen over the coming decade. Chip design relies on electronic design automation (EDA) tools. These tools help engineers translate high-level specifications into detailed circuit layouts, then verify that designs meet performance, power and reliability requirements. While EDA tools are highly capable, they are typically driven through long sequences of manual steps. Engineers configure tools, run simulations, review results and adjust parameters repeatedly until designs are ready for signoff. This approach is proven, but it is time-intensive. It requires sustained expert attention and significant computing resources. As designs grow more complex, late-stage changes become harder to manage and more costly to correct. The solution Goal-driven assistance within existing design environments AMD and Synopsys are working together to introduce AI-supported workflows that assist engineers throughout the chip design process. Rather than replacing existing EDA tools, the approach embeds intelligent systems directly into the environments engineers already use. These systems interpret design goals, coordinate tool runs and evaluate intermediate results. Engineers remain responsible for decisions, while routine execution tasks are handled automatically. By shifting from manual tool operation to goal-driven workflows, teams can reduce repetitive effort. This allows engineers to focus on architectural choices and trade-offs while maintaining full visibility into the design process. How it works Learning from real design activity The system uses components capable of reasoning about design objectives and planning sequences of actions across the toolchain. It learns from historical design data as well as from interactions with engineers during active projects. Training draws on years of internal design experience from AMD and the capabilities of Synopsys’ EDA platforms. Over time, the system can recognize recurring design patterns and flag potential issues earlier in the flow. It can also help engineers explore alternative design options more efficiently. This support improves iteration speed while keeping human designers in control of key decisions. How it differs from conventional approaches Reducing manual orchestration Traditional chip design workflows require engineers to manually orchestrate each step of the EDA process. Adjustments are often reactive, based on results that appear late in the cycle. AMD and Synopsys’ approach reduces reliance on manual coordination by allowing AI-supported systems to manage execution across tools. Engineers set objectives and constraints, then review and refine outcomes as the system progresses through the workflow. This helps identify potential issues earlier and reduces the likelihood of disruptive late-stage changes. Impact Productivity across design and verification stages approximately doubled Design exploration expanded by 25%, allowing teams to evaluate a wider range of options Overall design costs reduced by a factor of five Time to signoff reduced by 50%, with fewer late-cycle changes Faster design cycles contributing to improved reliability and fewer defects Why it matters Making expert time go further As chip complexity continues to increase, design productivity depends on how effectively expert engineers can focus their attention. The collaboration between AMD and Synopsys shows how AI-supported workflows can help teams make better use of limited expertise without changing established tools or processes. By supporting engineers throughout the design flow, this approach helps shorten development cycles and manage complexity in an industry where both time and talent are increasingly constrained. 
USA | Technology
Reinforcement learning, generative AI and agentic AI for chip design
Synopsys, AMD
Cohort: 2 Location: Wuhan, China Application Title: Automated Cytology and Telepathology for Population Cervical Cancer Screening Application Subtitle: A digital cytology and telepathology system that automates cell analysis and connects clinics to remote specialists, enabling large-scale cervical cancer screening where pathologist capacity is limited Team Members: Lan Wang, Xiaorong Sun, Baochuan Pang, Cheng Li and Jian Wang The challenge Cervical cancer remains one of the leading causes of cancer-related deaths among women in low- and middle-income countries, despite being largely preventable through early detection and treatment. A major barrier is the global shortage of trained pathologists. In China, for example, there is an estimated shortage of 100,000 pathologists. This shortage makes it difficult to scale cytology-based screening, which remains one of the most cost-effective and clinically validated methods for cervical cancer detection. Although HPV DNA testing is often promoted as an alternative, it is significantly more expensive and difficult to sustain at population scale in resource-constrained settings. Existing efforts to expand screening rely on time-consuming and unstandardized manual cytology review. These conditions are rarely met in rural or underserved areas. As a result, screening coverage remains uneven, diagnoses are often delayed, and health disparities persist. The solution AI-assisted cytology at population scale Landing Med addresses these constraints with an AI-powered cytology platform that automates cell analysis and connects clinics to remote pathologists through a cloud-enabled screening workflow. Cytology involves examining cells under a microscope to detect early signs of disease. Landing Med’s platform digitizes this process by combining slide scanning, deep learning–based image analysis and telepathology into a single end-to-end service. The system is designed to operate without requiring extensive local infrastructure, making it suitable for deployment across diverse clinical settings. To date, the platform processes approximately 190 terabytes of medical imaging data, supporting high-resolution cytology analysis across large populations. How it works From sample collection to clinical review Partner clinics collect and prepare samples locally or sent to a Landing Med lab. Landing Med’s AI screening system integrates distributed local sample preparation with centralized, AI-driven diagnostic support and analysis, enabling scalable, standardized, and governable cancer screening at population scale.The system employs clinically validated AI models to perform multi-level risk stratification and pre-screening of digital pathology slides. These models prioritize cases based on diagnostic risk, supporting a structured human–AI collaboration in which cytotechnologists and pathologists focus their expertise where it is most needed. The performance, updates, and clinical alignment of the models are centrally governed through a unified framework, supporting transparency, accountability, and sustained trust across large-scale screening environments.Supported by an elastic and scalable computing infrastructure, the system is able to dynamically scale to meet the demands of large-scale public health screening programs. Through real-time, intelligent quality control and analytical monitoring, the system ensures consistent diagnostic performance while optimizing the allocation of clinical, technical, and human resources, enabling high-quality screening to be delivered efficiently and sustainably at population scale. How it differs from conventional approaches Reducing dependence on manual review Conventional cytology screening depends heavily on manual slide review and limited specialist capacity. Landing Med’s approach uses AI to triage large volumes of samples, reducing bottlenecks without removing clinical oversight. By integrating AI-driven analysis, telepathology and deployment infrastructure into a single platform, the system supports coordinated screening across clinics, hospitals and laboratories. This reduces fragmentation and allows screening programmes to scale more consistently. The model has already been deployed across more than 2,000 hospitals and clinics in China and can be adapted to other low-resource settings. Impact More than 13 million AI-assisted screenings conducted since 2017 91% coverage across Chinese provincesFivefold increase in screening productivity (from 12 to 60 samples per hour) 90.1% sensitivity and 94.8% specificity for CIN2+ detection Up to 45,000 screenings per day at peak capacity 10% random rechecks of negative cases to support quality control  Why it matters Expanding access to early detection Landing Med demonstrates how AI can support population-based cervical cancer screening in settings where specialist capacity is limited.By reducing reliance on manual review while maintaining clinical oversight, the platform enables earlier detection at scale and supports more consistent screening coverage. This approach addresses practical barriers to prevention and contributes to improved outcomes for women in underserved regions.
People's Republic of China | Health
Automated Cytology and Telepathology for Population Cervical Cancer
Landing Med
Cohort: 1 Locations: Amstelveen, Netherlands & Walldorf, Germany Application Title: AI copilot for ERP acceleration and cloud transformation Application Subtitle: Team MembersKMPG: Hendrick ThörnerSAP: Philipp Herzig (Executive Sponsor), Sachin Kaura, Natalie Han, Joachim von Goetz, Christian Karaschewitz, Caroline NgoThe challenge Many large organisations rely on enterprise resource planning (ERP) systems to manage finance, supply chains and core operations. A significant share of these systems were built on on-premise infrastructure and are now being migrated to the cloud. These migrations are technically complex and often span several years. ERP transformation projects require consultants and internal teams to navigate extensive technical documentation, system-specific rules and legacy custom code. For SAP environments, this often involves working with ABAP, a programming language used to customise business processes. Understanding what existing code does, whether it remains compliant in the cloud, and how it should be adapted takes substantial time. Traditionally, teams rely on manual searches through documentation, static guides and community forums. This slows progress, increases the likelihood of errors and leads to rework when incorrect assumptions are discovered late in the project. The solution An AI copilot grounded in ERP expertise KPMG partnered early on to leverage SAP’s AI capabilities via SAP Joule for Consultants, which is designed to support ERP migration teams during cloud transformation projects.The copilot is built on a vast body of trusted SAP documentation and best practices to answer consulting questions, explain ABAP code, and suggest configurations helping practitioners move faster, with less error and work with greater confidence. It draws from SAP documentation of more than 100 million pages, along with curated best-practice materials and validated implementation guidance (up to 12 Terabytes of content - 25 million documents and growing). Instead of acting as a general chatbot, the system is tailored to answer questions specific to ERP configuration and migration work. KPMG has additionally grounded their SAP Joule for Consultants instance with KPMG’s proprietary knowledge from past migration projects. Consultants can ask questions in plain language and receive natural language responses that are based on and directly linked to authoritative sources.How it works Retrieval-augmented guidance during implementation The copilot uses a retrieval-augmented architecture that connects user questions to curated SAP knowledge. When a query is submitted, the system retrieves relevant content and uses language models to explain it in context. For developers, the copilot can analyse ABAP code and explain its purpose, structure and dependencies. It can also suggest code changes or generate new code when appropriate. This reduces the need for manual interpretation and helps teams assess whether customisations should be migrated, modified or retired. The system operates within existing SAP project environments and supports human review, ensuring that final decisions remain with experienced consultants. How it differs from conventional approaches Replacing manual search with contextual assistance Conventional ERP migrations depend heavily on keyword searches across large document repositories and individual expertise. Information is fragmented and often outdated, making it difficult to apply consistently across projects. The copilot consolidates prior learning and existing knowhow into a single interactive system. Rather than searching across multiple sources, teams receive contextual guidance at the point of work. Responses are grounded in validated content, reducing the risk of incorrect configuration or non-compliant code. This approach shortens feedback loops without removing professional judgment. Impact Up to 1.5 hours saved per consultant per day Reduce average project duration from 18 to 15.5 months Up to 18% faster execution, translating to approximately USD 1.3 million in potential savings for a 2,000-employee program  Reduced rework and improved quality through expert-aligned guidance based on current SAP best practices  Faster understanding of custom ABAP code, reducing interpretation time and accelerating remediationWhy it matters Reducing friction in large-scale ERP change ERP migrations are among the most complex transformation programmes organisations undertake. The collaboration between KPMG and SAP shows how AI can support these efforts by reducing time spent on documentation and code interpretation, while building on both's organizations strengths.By embedding expertise directly into daily workflows, the copilot helps teams move faster while maintaining accuracy and compliance, supporting more predictable outcomes in long-running transformation projects. 
Germany, Netherlands | Technology
AI copilot for ERP acceleration and cloud transformation
SAP, KPMG
Cohort: 2 Location: Ningde, China Application Title: Physics-informed and agentic AI platform for EV battery cell design Team Members: Prof. Ni Jun, Dr. Jin Wenjing, Dr. Huang Yunlong, Dr. Jin Haizu, Song Shutao, and Ying YuanyuanThe challenge Rapid growth in electric vehicle adoption has significantly shortened battery development timelines. What once took several years now often needs to be completed in roughly 18 months. This places sustained pressure on engineering teams to deliver improvements in performance and cost efficiency within increasingly tight schedules. Historically, battery cell design has relied on expert intuition supported by manual trial-and-error. Engineers managed thousands of interdependent parameters through long prototyping cycles, adjusting chemistry, structure and performance characteristics incrementally. The process worked, but it was slow, resource-intensive and difficult to scale. As battery systems became more complex, the limits of human-led iteration became more pronounced. More recently, AI-assisted models trained on historical data have shortened design cycles from weeks to days. While this marked a clear improvement over manual workflows, these approaches still depend on sequential optimization steps and tightly structured inputs. In practice, that constraint limits their ability to adapt to the complexity of real-world battery design. The solution An AI-powered cell design platform grounded in physics CATL has developed an AI-powered battery cell design platform that integrates physics-based electrochemical modelling with machine learning to automate and accelerate key parts of the design process. Physics-informed models incorporate established electrochemical principles directly into the learning process, helping ensure that predictions remain consistent with known physical constraints. These models are combined with machine-learning techniques to evaluate large design spaces efficiently. The platform runs on an on-premises private cloud, providing secure high-performance computing while maintaining full data sovereignty. It processes more than 50 million data records, spanning time-series measurements, text, images and graph data. Structured and unstructured inputs are combined with synthetic data generation, allowing training datasets to expand and improve over time. Across the organization, the platform is used by more than 4,200 employees across 15 departments, including over 1,200 core engineers. What previously took weeks of iteration can now be completed in minutes. How it works From manual optimization to autonomous design loops At the heart of the platform is a combination of physics-informed machine learning and agentic AI. Agentic AI refers to systems that can plan, evaluate and refine actions autonomously within defined constraints. Within CATL’s platform, this creates what the company describes as a “digital engineer” workflow. The system autonomously generates design candidates, evaluates them against electrochemical constraints, and refines solutions through continuous iteration. Instead of relying on human-led optimization loops, the platform operates as a closed system where data, physics and computation reinforce each other. By embedding domain knowledge directly into the model, CATL ensures that automation does not come at the expense of scientific rigor. How it differs from conventional approaches Reducing dependence on manual evaluation Conventional battery design workflows rely heavily on sequential testing and expert review. CATL’s platform reduces this dependence by enabling data-driven evaluation across a much broader design space. Design decisions are no longer constrained by the pace of manual evaluation. Instead, engineers can explore alternatives more quickly and focus their attention on validation and decision-making, rather than repeated parameter tuning. Why it matters Supporting faster and more reliable battery development By transforming battery cell design into a fully data-driven process, CATL is changing how next-generation EV batteries are developed. The platform shortens development timelines, improves predictive accuracy and allows innovation to scale without relying on ever-growing engineering teams. In doing so, it supports faster progress in battery performance while reducing cost and complexity, a critical step in powering the global transition to sustainable mobility. 
People's Republic of China | Energy Technologies
Physics-informed and agentic AI platform for EV battery cell design
Contemporary Amperex Technology (CATL)
Cohort: 2 Location: Beijing, China Application Title: City-scale AI orchestration system for efficient and low-carbon grid dispatch and virtual power plant operations Team Members: Wei Xie (executive sponsor), Guanxun Diao, Yuezhong Tang, Jun Zhu, Ling Luo, Xiao LuoThe challenge As one of the world’s largest and most complex megacities, Shanghai faces growing challenges in managing its power grid. Rising electricity demand, carbon reduction targets and the increasing share of variable renewable energy all add to operational complexity. The grid serves more than 1.7 million corporate users and 10.6 million residential users. Balancing supply and demand requires continuous coordination across generation, transmission, load and storage. Traditional forecasting and regulation systems struggle to manage this complexity in real time. Manual coordination is too slow for the second-level decisions in response to respond to renewable fluctuations, while expanding thermal generation or storage capacity entails high costs and additional carbon impact. The solution A unified AI platform for city-scale grid orchestration To address these constraints, the State Grid Corporation of China developed a city-scale AI orchestration platform that integrates forecasting, trading, regulation and settlement into a single operational system. The platform is built on the Guangming Large Model, a proprietary multimodal large language model trained on more than 200 terabytes of power system data and a domain-specific token library of 3.5 million energy-related terms. It continuously predicts and optimizes power flows across Shanghai’s grid using real-time data from edge gateways, enterprise systems and grid sensors. This provides detailed visibility across generation, transmission and consumption, enabling more coordinated management of distributed energy resources. How it works Real-time prediction and adaptive control The system combines spatiotemporal modelling, Model Predictive Control (MPC) and NSGA-II optimisation to support adaptive control with second-level response times. It operates on an on-premises architecture with edge–cloud coordination, allowing real-time orchestration of thousands of distributed energy devices. These include air conditioners, electric vehicle chargers and industrial equipment, which can be coordinated dynamically while maintaining user comfort and grid stability. The platform uses multiple intelligent agents to manage forecasting, trading, regulation and settlement within a unified workflow, supported by human-in-the-loop oversight through operational dashboards. This approach reduces reliance on manual balancing and enables more consistent responses to rapidly changing grid conditions. How it differs from conventional approaches From isolated tools to integrated orchestration Conventional grid management systems typically address forecasting, regulation and trading as separate functions. State Grid’s platform brings these capabilities together, allowing decisions to be coordinated rather than sequential. Instead of reacting to fluctuations after they occur, the system anticipates changes in demand and supply and adjusts control strategies accordingly. This architecture has enabled applications such as reactive power demand response in Shanghai’s Lin-gang industrial zone, helping stabilize voltage levels without requiring additional infrastructure investment. Impact 2.03 GW of dispatchable capacity coordinated across 49 virtual power plant operators and more than 4,000 users More than RMB 8 billion saved through avoided generation and transmission investments 1.236 billion kWh of clean energy consumed annually, with 510,000 tonnes of CO₂ reduced per year 99.999% voltage compliance and 99.9983% grid reliability Improvements of 12.5% in forecasting accuracy, 72% in regulation response, and 51% in settlement efficiency Over 15,000 active users across grid operations and virtual power plant teams Why it matters Managing grid complexity in renewable-heavy systems The State Grid Corporation of China’s platform illustrates how AI-based orchestration can support power system operations in large, renewable-heavy urban environments. By integrating forecasting, control and settlement within a single system, the platform helps maintain grid stability while reducing the need for costly infrastructure expansion. This approach supports more efficient use of existing assets as cities work to meet growing demand and decarbonisation goals. 
People's Republic of China | Energy
City-scale AI orchestration system for efficient and low-carbon grid dispatch and virtual power plant operations
State Grid Corporation of China
Cohort: 2 Location: Pune, India Application Title: Sovereign multilingual framework inclusive digital public services Application Subtitle: A multilingual language framework trained in local languages and dialects, enabling culturally grounded digital services while retaining control over data and infrastructure Team Members: Mohit Joshi, Sham Arora, Nikhil Malhotra  The challenge Digital public services increasingly rely on language-based interfaces. Citizens interact with government portals, banking systems and healthcare services through chatbots, voice systems and automated forms. Most of these systems are powered by large language models trained primarily in English and a small number of high-resource languages. In countries with high linguistic diversity, this creates a barrier. Many users speak national languages such as Hindi or Bahasa, often alongside regional dialects that are poorly represented in global datasets. As a result, automated systems struggle with intent recognition, translation accuracy and cultural context. This limits access to digital services and raises concerns about data sovereignty when sensitive interactions depend on externally hosted models. Conventional approaches attempt to adapt global models through translation layers or lightweight fine-tuning. These methods improve surface-level performance but often fail to capture linguistic nuance, especially for dialects and informal speech. The solution A sovereign language model built for low-resource languages Tech Mahindra developed a sovereign multilingual AI framework designed specifically for national languages and regional dialects. Rather than adapting a global model, the system is trained from the ground up using data sourced from national archives, public literature and domain-specific corpora in sectors such as healthcare and finance. Training involves reinforcement learning with native speakers, ensuring that responses reflect local usage and cultural context. The models are deployed on sovereign infrastructure, allowing governments and institutions to retain control over data and system behavior. This design supports use in public-facing services where privacy and governance requirements are strict. How it works Combining language understanding with local knowledge At its core, the system uses transformer-based language models adapted for Hindi, Bahasa and selected regional dialects such as Bhojpuri and Javanese. These models support both text and speech, making them suitable for conversational interfaces. During inference, retrieval mechanisms pull verified information from curated national repositories. This helps ground responses in authoritative sources rather than relying solely on model-generated text. Training and deployment are supported by national high-performance computing environments, enabling scaling without transferring data to external platforms. Built-in safeguards address content moderation and error detection, allowing deployment in high-stakes contexts such as citizen services and education. How it differs from conventional approaches Moving beyond translation-based adaptation Global language models typically perform well in English but lose accuracy when applied to low-resource languages. Performance drops further when users rely on dialects or informal speech. Tech Mahindra’s approach avoids this limitation by treating linguistic diversity as a design requirement rather than an edge case. Models are trained directly in target languages and validated by native speakers. This improves intent recognition and reduces translation errors without adding intermediary layers. Because the system runs on sovereign infrastructure, it also addresses concerns around data residency and long-term autonomy. Impact Conversational accuracy increased to 92% in Hindi and Bahasa, compared with a global benchmark below 70% Translation errors reduced by 35% Approximately 3.8 million queries per month supported across public services, banking and healthcare Active use across government platforms and institutional deployments Why it matters Improving access to digital services in multilingual societies Language remains a primary interface between people and institutions. Tech Mahindra’s framework shows how AI systems can be designed to reflect linguistic reality rather than forcing users to adapt to global defaults. By supporting national languages and regional speech patterns, the approach helps widen access to digital services while maintaining local control over data and governance. This becomes increasingly important as AI-driven interfaces play a larger role in public life. 
India | Government
Sovereign multilingual framework inclusive digital public services
Tech Mahindra

MINDS Champions

Loading

MINDS

A World Economic Forum Initiative
© 2026 All rights reserved
Privacy Policy & Terms of Service