← sources

Small Indian team building AI infra, looking for 3-4 pilot teams

source: reddit (reddit:IndianStartups)published 2026-10-20 10:10ingested 2026-10-21 02:37

retained text

Hey all, we're a small team from India building MLJunction. We started because using multiple LLMs in one product was more painful than we expected.

Swapping models broke things. Every model has its own rules for temperature, thinking tokens and tool calls, and we kept losing days to provider docs. We tried global routers and liked them, but we still hit those quirks, and I suspect conversations hopping between providers cost us prompt-cache savings.

So we built the layer we wanted:

1. You write one request and it's translated for whichever model it lands on. The response lists every change it made.
2. Long history is trimmed to fit, but your system prompt and the latest user message are never dropped, and broken tool messages get repaired.
3. A conversation stays on the provider that holds its cache, so repeated context bills at the cached rate.
4. If a provider fails, the request moves to the next route automatically.
5. You can name a pool of models and let them compete on your own outcomes (ticket resolved, test passed). A challenger earns traffic in stages, and every change can be rolled back.

You can also pay in rupees (UPI, UPI AutoPay) if that helps, but honestly that's the smallest part of it.

We're early and I'd rather learn than sell. If your team runs LLMs in production, I'll set it up with you personally. All I ask for is blunt feedback. And if you've tried something like this before, I'd love to know what put you off.

document fetched 2026-10-21 00:26 · canonical https://reddit.com/r/indianstartups/comments/1xao7y1/small_indian_team_building_ai_infra_looking_for