[ A lab for curious minds ]

Intelligence by
deduction.

We believe superior reasoning emerges from superior environments. Through more rigorous benchmarks and open training ecosystems, we are setting a new standard for how smarter AI is measured and built.

01 / Research tracks

We split our work into four research domains and four methods. Benchmarks expose gaps. Data, environments, and training close them.

Research domains

[ D-01 ] Reasoning

Reasoning Models

Harder STEM problems where the frontier still fails.

[ D-02 ] Embodied

Robotics & Physical AI

Models that see, reason, and act on physical objects. We focus on unstructured environments.

[ D-03 ] Speech

Indic Speech & Language

Speech recognition, synthesis, and language understanding for Indian languages.

[ D-04 ] Software

Coding / SWE

Models that navigate full repos, use tools, interpret execution feedback, and produce verifiable fixes.

Core methodology

  1. 01

    [ M-01 ] Evaluation

    Benchmarks

    Evals that resist contamination and stay hard on purpose. We want to know where models stop working, not how close they are to a ceiling everyone already hit.

  2. 02

    [ M-02 ] Data

    Synthetic Data

    We generate training data around the specific failure modes our benchmarks find.

  3. 03

    [ M-03 ] Environments

    RL Environments

    Deterministic environments for coding in low-resource languages and domains where the frontier still breaks.

  4. 04

    [ M-04 ] Training

    Training

    RL and post-training on open-weight models, using the data we produce and the environments we build around them.

02 / Projects

What we've shipped so far.

D-01/Reasoning Models

Aryabhata 1

Open source · Qwen 2.5 7B

[ NeurIPS 2025 MATH-AI Workshop ] 7B reasoning model for JEE Main Math, built on Qwen 2.5 with on-policy curriculum SFT and GRPO

Read project log →

03 / Research grants

We fund researchers who build benchmarks that expose real weaknesses in frontier models. We co-develop the full stack from eval design through training, and provide compute.

Send a one-page proposal with your problem statement, approach, timeline, and compute needs. We review on a rolling basis.

Apply for a grant