Navital Labs is an independent AI lab in Delhi.
We build Navital Gym: open environments where AI agents practise real Indian tasks, graded by written rules.
Navital Gym is live on GitHub. Results are updated as runs finish; the full report is due 10 November 2026. See the leaderboard · Open the repository
Navital Gym is a set of five environments. In each one an agent works in a small simulated world and a program scores the result.
- Trip Planningmulti-city trips under budget, rail quota, closing days and food needs.Live
- UPI Deskpayments, limits, failed payments and scam requests.Live
- GST Deskinvoices, input credit, due dates and late fees.Live
- Form Deskidentity forms, addresses and Indian number and date formats.Live
- Bhasha Support Deskshop support in Hindi and Hinglish, plus an experimental Tamil, Bengali and Marathi split.Live
Many public environments and evaluations we studied are set in Western workflows. Indian tasks differ in ways an agent has to handle: rail classes, quotas and waitlists; UPI limits and failed-payment rules; GST slabs and place-of-supply rules; PIN codes and lakh-crore numbers; customers writing in Hindi, Tamil, Bengali or Marathi. We are building the missing environments, and will publish them for others to run.
An environment has tasks, tools and a grader. The agent sees the request and the tools. The grader applies written rules to what the agent did: which payment was made, which tax applied, which fields were filled. A language model does not decide the core score. Some tasks have no valid answer, and the correct response is to say so and name the rule that blocks it.
We also build products, because real work shows what agents get wrong.
Research teams, evaluation groups and agent builders who need environments for Indian workflows. Native speakers of Indian languages who can review task text. Contributors who find a wrong rule.
Tell us the task and what a correct result looks like. We will say whether we can build it and what we would need.