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.