Tech Mahindra's Project Indus, an initiative to build a foundational large language model rooted in Indic languages, now handles 3.8 million queries a month for banking, healthcare and government organisations, the company said in February 2026. The project was recognised by the World Economic Forum's MINDS programme at its 56th Annual Meeting for advancing linguistic and digital equity through AI.
The project started from an accessibility gap. India has 22 official languages and more than 1,600 dialects, with only 10 to 20% of the population speaking English, yet many public services in healthcare, banking and government are delivered through chatbots and IVR systems built on models trained mainly in English and a few Western languages. Such models often fail to recognise intent and cultural context in regional languages and translate inaccurately.
Rather than retrofitting a global model with a translation layer, Tech Mahindra built its own model with linguistic diversity as a design principle. It was trained on datasets drawn from open national databases, literature, archives and user contributions, and refined by native speakers. The model supports both speech and text and can be adapted to any conversational interface, and because it is the company's own model, enterprises can keep full control over their data and the model's behaviour. Project Indus also aims to preserve and digitise Indic languages and dialects with few digital records and to produce localised, industry-specific Indic models, with partnerships planned with telecoms operators, hyperscalers and device makers to offer end-to-end LLM solutions.
Tech Mahindra says the model reaches 92% accuracy in Hindi, compared with a global benchmark of 70%.
Details
- City
- Pune
- Organization
- Tech Mahindra
- Continent
- Asia
- Country
- India
- Category