Five tracks, and a gate.

6–10 contributors

World Models and JEPA

Joint Embedding Predictive Architectures point past purely generative AI toward sample-efficient learning, grounded reasoning and world models for embodied agents. For a research programme with limited compute, an architecture family whose central claim is sample efficiency is not a fashionable choice — it is the practical one.

Kill criteria. If replication does not reach published baselines within two quarters, the track narrows to the Indic low-resource question only, or closes.

Read the track brief →

6–10 contributors

Tokenomics — Valuing AI Token Usage

As language models become infrastructure, the unit economics of inference remain poorly understood. Pricing, fair use and per-token value attribution are open problems with genuine policy weight — and they are a live blocker on enterprise adoption, which makes them squarely an SME problem.

Kill criteria. If the attribution toolkit cannot reproduce a finance team's own cost model on a real workload, the framework is wrong and the track closes rather than being defended.

Read the track brief →

4–8 contributors

Maharashtra AI Policy

Maharashtra is writing AI policy while the industrial base it applies to — Pune's manufacturing and services SMEs — has almost no representation in the drafting. Policy written without that input tends to fund capability that already exists in large firms and misses the constraints an SME actually faces. The window for evidence to be useful is while a document is being drafted, not after it is published.

Kill criteria. If the policy process closes without our submission being read, or the work becomes advocacy without evidence behind it, this closes rather than continuing as a relationship-building exercise. Influence that cannot be pointed at is not a deliverable.

Read the track brief →