Case Study - Turn any video into searchable, speaker-labelled knowledge
An AI transcription and video-understanding pipeline: deep transcripts synchronized to an embedded player, with exports and automatic channel processing.
- Project
- spectralTranscript
- Year
- Service
- AI pipeline, full-stack product
Overview
Video is where knowledge goes to become unsearchable. spectralTranscript turns any YouTube video into structured, speaker-labelled text you can read, search, and export — with the transcript synchronized to an embedded player so you can jump straight from a sentence to the moment it was said.
Under the hood it's a transcription pipeline: ingest via the YouTube Data API, deep transcription with speaker labelling, storage on Cloudflare R2, and an automatic processing queue so a channel's new uploads become searchable without anyone pressing a button.
What we did
- TanStack Start (SSR)
- Hono + oRPC API
- PostgreSQL + Drizzle
- AI transcription & speaker labelling
- Cloudflare R2 storage (Alchemy)
- Docker / Nixpacks deployment
The product is simple to describe — video in, knowledge out — which is exactly why the pipeline behind it has to be boring, observable and automatic.
Engineering notes
spectralTranscript is also our reference implementation for TanStack Start in production: server-rendered routes, typed server functions, and a queue-driven backend that scales the expensive work (transcription) independently from the reading experience.