Transcend
Near real-time speech transcription and translation across 30 languages — built during NCST's virtual on-the-job training, where I hardened the backend and owned the WebSocket layer.
The problem
Live speech has to be captured, transcribed, and translated fast enough to feel instant — across many languages and unreliable network conditions. Naive implementations drop connections, double-bill translation APIs on retries, and render half-formed text that's hard to read.
Approach
I focused on the reliability layer: per-socket rate limiting to protect the service, an idempotency cache that prevents duplicate translation-API charges on retries, and request retries with exponential backoff and timeouts. I integrated the real-time WebSocket pipeline end to end, added sentence-boundary detection so captions stream cleanly line by line, and wrapped the client in an error boundary with automatic socket reconnection.
My role
Top contributor on the team. I owned backend hardening and the front-end-to-back-end WebSocket integration, plus the transcription pipeline's segmentation and resilience work.