Make AI in India, for India
A small team building AI models that millions of people will use.
Come and pretrain AI models from scratch
Find your team
Our teams work from Mumbai, Hyderabad, Pune and Delhi. Most roles are in Mumbai, with hybrid working.
Train, serve and ship our AI models.
Generative AI, ML research, data, platform, full-stack and agentic AI
- Pretraining and post-training text, speech and vision models
- Data pipelines and tokenisers for Indian scripts
- Serving models and building agentic applications on them
Make sure our models speak India's languages well.
Linguists and language data specialists
- Checking our models' answers across 22 Indian languages
- Building benchmarks and evaluation data
- Curating language data with our engineers
Take our models from the lab to the people who use them.
Product, program management, partnerships, finance and people
- Product for the applications built on our models
- Partnerships and deployments with hospitals, newsrooms and governments
- Programmes, finance and people operations
Learn by working on real models.
From undergraduates to postdoctoral researchers
- Real models alongside the team, not side projects
- PhD and M.Tech fellowships across the consortium
- A path to a full-time role
What we value
We never wrote these down in a meeting. They grew out of the work. Turn a card to read where each came from.
Whether it works, not where it ranks
A model has to work for the people who use it, not top someone else's leaderboard.
Standard benchmarks rarely tested Indian law, farming or finance, so the team built BhashaBench and put every model through it, ours included.
Get in the room
We test with real users early, in deep pilots, before we think about scale.
Krishi Sathi was trained to help farmers grow crops. One of its first users, a farmer near Nashik, asked how to store his onions.
Built for how India lives
Language is only the start. Our models should get Indian names, mixed scripts and local customs right.
A teammate pointed out that global AI models tell you to lock the door on a stranger. Where he grew up, you offer them a glass of water.
If it can be done, we can do it
We build from scratch, with our own data and tokenisers, rather than adapting someone else's model.
When we were starting out, we were told publicly that India was too late to build its own foundation AI models. The answer has been the work.
Frugal by design
AI only matters if India can afford it, so we design for low cost to train and to serve.
Frugal scaling has guided our research from the start. Param2 runs only about 2.4 billion of its 17 billion parameters for each word.
Open by default
We publish weights, papers, benchmarks and recipes so others can check our work and build on it.
Hospitals and governments told us they need to run AI on their own terms, with their own data. Open weights make that possible.
Who else can we invite?
We would rather bring partners in than compete. People who know a field build its best applications.
Nine institutions came together to build BharatGen, and doctors, editors and officials now build on our models.
The hard problems are here
AI for someone whose first computer is a phone and a voice note is frontier work, and it is happening here.
The question our CEO asks about every young engineer: not whether they should leave, but whether what we are building is worth staying for.
Build the builders
Teaching is part of the job. Interns work on real models alongside the team, not on side projects.
More than 190 interns so far, from undergraduates to postdoctoral researchers.
"It's not just about building the models. It's not just about building the applications. It's about building the builders."
What it is like here
Work that ships
Models you train go out on Hugging Face and into hospitals, newsrooms and government offices.
Learn from the best in the country
Work with researchers from nine top institutions, and present at top conferences.
Ownership from day one
Small teams, real responsibility.
Mission first
People who join us care about who the work is for.
See our open roles
Every opening is posted on LinkedIn, with the details of the role and how to apply.