This is a build log.
The idea
Regional language support is the actual moat in the Indian AI market. English-first products serve maybe 200 million people. Indic-language products serve 1.4 billion. The gap is enormous and almost nobody is filling it well.
CrewSpace's premise: a non-technical user — a small business owner, a teacher, a local NGO worker — can build an AI agent workflow in their own language. No code. No English required. Drag, drop, describe in Assamese or Hindi or Bengali, deploy.
The stack we chose
We went Google-aligned deliberately. Google Solution Challenge was on our minds, and aligning with their stack gave us better free tier limits and a coherent story.
- Frontend: Angular (I'm comfortable, fast)
- Backend: FastAPI on Cloud Run
- AI: Gemini 1.5 Flash for the agent logic, Sarvam AI for Indic language processing
- Database: Firestore
- Auth: Firebase Auth
The decisions that saved us
Fake the hard parts in the demo. The full agent execution pipeline wasn't complete at hour 28. But the demo didn't need it to be complete — it needed it to look complete. We hardcoded the happy path for the presentation, clearly labelled as a demo flow, and spent the remaining time making that one flow bulletproof.
Cut scope ruthlessly. The original plan had five agent types. We shipped one, done properly. The judges never knew there were meant to be four more.
Sarvam AI for transliteration, not translation. We initially tried to translate the entire UI. It was slow and the quality was inconsistent. Switching to transliteration — keeping the structure English, rendering text in the user's script — was a 2AM decision that saved the product.
The decisions that cost us
We over-engineered the agent execution model in the first 12 hours. I wanted a proper queue system with retry logic and status streaming. We didn't need that for a hackathon. That 6 hours cost us polish elsewhere.
What CrewSpace became after
After winning, we kept building. The pricing model, the template marketplace, the team plan tiers — none of that existed at the hackathon. The hackathon version was a proof of concept. The real product is being built properly now, with actual architecture decisions instead of 3AM ones.