Capstrix — a voice-data marketplace

Capstrix is a product I’m bootstrapping end to end. The bet: the most valuable speech data for training voice models — natural, code-switched, conversational audio in the languages people actually mix when they talk (Hinglish, Tanglish, Banglish) — barely exists. Capstrix collects it properly: consented, native-speaker, transcribed, and speaker-verified.

What I’ve built

Because I’m the only engineer, the whole thing is designed to be small, typed, and hard to break:

• A mobile recording app with a focused capture flow and passwordless sign-in, distributed through TestFlight.
• A typed API on the edge, live on its own domain, with every client surface routing through one shared contract.
• An admin console for managing contributors, roles, and review.
• An automated audio-QA pipeline that runs every recording through quality checks before it’s accepted.

The architecture I’m proud of

One shared, typed contract is the single source of truth across the mobile app, the admin console, and the API — change a schema in one place and every surface either updates or fails to compile. For a solo project, that discipline is the difference between moving fast and drowning. The whole system runs on a serverless edge stack, so there’s almost no infrastructure to babysit.