Skip to content
Navital Labs

Labs

Navital Labs builds environments where AI agents practise real Indian tasks, with automatic grading.

Five environments, live on GitHub, starting with planning a trip in India under real constraints.

Private evaluation of your agent: $750 plus taxes ↓Read about Env #1 ↓Contact: dakshbathla@navitallabs.com

01 · WORK WITH US

Work with us

If you need an Indian workflow as an environment, tell us the task and what a correct result looks like. We reply with whether we can build it and what we would need.

Private evaluation

$750 per model or agent version, plus applicable taxes

Want to know how your model or agent handles Indian tasks before you ship it? We run it on Navital Gym's private held-out tasks in all five worlds (six task sets: Form Desk also has a harder multi-document set). The held-out tasks have never been published, and your scores won't be either: you get a private written report. Paying for an evaluation does not change how any public score is graded or listed.

  • 181 held-out tasks, about 30 per set: trip planning, UPI, GST, Form Desk (standard and multi-document) and support chats in Hindi and Hinglish.
  • Graded by the same written rules as the public leaderboard. On every held-out set, each scripted baseline except the reference solver scores under 25%.
  • A private written report within 5 working days: pass rates per world, the main failure types, and fixes to try.
  • What we need from you: API access or an endpoint, and you pay your own model usage. Happy to sign your NDA.
  • One version per evaluation; a new version is a new evaluation at the same price. Invoiced 100% upfront, paid by bank transfer or Wise.

Email us about a private evaluation

Pilot

From $2,500

A focused slice of 20-30 tasks built around your use case, with grader, reference solutions and a short report. About two weeks.

Custom environments for AI labs

Private task sets from $5,000

The $2,500 Pilot is a 20-30 task slice around one team's use case, with a short report; this is private task sets and graders that an AI lab keeps, for training or evaluation.

We build private reinforcement-learning environments and tasks set in Indian workflows: UPI, GST, KYC and forms, travel, and Indic-language support. A new world is scoped per project. Every task is marked by a deterministic, rule-based grader, never an AI judge. You receive private task sets and graders in Navital Gym's open format, and an exclusive set is available. A first set is a small private set in one world; larger sets are scoped per project.

  • Deterministic, rule-based graders: no AI judge, the same answer always gets the same score.
  • Each task set ships with an oracle solver, leak checks, and a degenerate-agent gate.
  • Private task sets and graders in Navital Gym's open format; exclusive sets available.
  • Private task sets from $5,000 (one world, small private set); larger sets scoped per project.

Email us about custom environments

Email dakshbathla@navitallabs.com

Prices are in US dollars. The pilot price is the starting point; the final scope is agreed in writing before work begins.

Navital Gym report

02 · WHAT THE LAB DOES

An environment is a small, self-contained world with tasks and a scoring rule. An agent acts inside it, and the score says how well it did. Researchers use environments to measure agents and to train them.

We build these environments around work that happens in India. The first covers trip planning: trains with quota and waitlist rules, closing days and seasons. The set also covers UPI payments, GST invoicing, forms and addresses, and customer support in Hindi and Hinglish, with an experimental split in three more languages. Each environment has fixed data, a fixed set of tasks, and a grader that applies written rules, not a model’s opinion.

03 · WHY INDIA

The public trip-planning and tool-use benchmarks we studied are set in Western workflows. Indian ones differ in ways an agent has to handle: rail classes and quotas, waitlists, festival and peak-season pricing, vegetarian and Jain food needs, tasks written in Hindi or mixed with English.

We are building the missing ones, and we will publish them for others to run.

04 · PROJECTS

Our projects are published together as Navital Gym, an open set of environments for Indian agent tasks. Live. Leaderboard · Repository

Status: Live

Trip Planning: Indian trip planning under real constraints

An agent receives a request such as a five-day trip across several cities with a budget, a group, food needs and a rest day. It searches a simulated rail, hotel and attractions world and returns a structured plan. The grader checks the plan against the rules: budget, dates, connections, seat availability, opening days, food and room limits. A request with no valid answer is part of the set; the correct response is to say so and name the rule that blocks it.

  • 55 task requests, most in English, some in Hindi and mixed Hindi-English.
  • Data is synthetic or from public sources. No customer data is used.
  • The rail, hotel and seasonal rules are a documented simplification of public Indian practice, not a live booking system.
Status: Live

UPI Desk

An agent handles payments and disputes in a simulated bank. It checks that the payee name matches, respects payment limits, deals with a payment that failed after the money left the account, and declines requests that look like scams. The grader checks which actions the agent took and which it should not have taken.

Status: Live

GST Desk

An agent works out GST for a small business: the tax split on an invoice, which tax applies to a sale, how much input credit can be claimed, and when a return was due and what late filing costs. The grader checks every number exactly.

Status: Live

Form Desk

An agent fills identity and bank forms from messy documents, reads Indian addresses and PIN codes, and understands amounts such as “सवा दो लाख” or “2.5 lakh” and Indian dates and financial years. The grader checks each field.

Status: Live

Bhasha Support Desk

An agent answers online-shop customers in Hindi or Hinglish (plus an experimental, machine-assisted Tamil, Bengali and Marathi split): returns, cancellations, damaged items and refunds. The grader checks the resulting order state, that nothing forbidden was done, and, for Hindi and Hinglish, that the reply is in the customer’s language.

All five use synthetic data and simplified rules. Names such as UPI, GST and Indian Railways describe the workflow; Navital Labs is not affiliated with NPCI, RBI, GSTN, CBIC, Indian Railways, IRCTC or any bank, and nothing here is financial, tax or travel advice.

Open source, with baseline scores for small, open-weight and commercial models. The code, tasks and every run file are on GitHub.

Explainer: What is an agent environment? →

Engineering notes from our products moved to /notes.

05 · OPEN SOURCE

Navital Gym is published openly, and others may run it, change it and report scores. Repository: github.com/Navitallabs/navital-gym.