A set of Go tools that monitor Twitch channels continuously, measure chat excitement and trigger clips at the right moment, then transcribe them with Whisper at zero API cost. The end goal: build a “context → clip” dataset to train a detection model.
Context
Analyzing video in real time is expensive. Chat, on the other hand, is free and reacts instantly: message velocity, emote density and certain keywords are excellent signals of a clip-worthy moment.
What I built
- twitch-auto-clipper: anonymous IRC chat reading, sliding-window “hype score”, Helix clip creation with cooldown, minute-by-minute stats in SQLite.
- twitch-clip-scraper: top clip collection, enrichment (game, VOD title), local or OpenAI Whisper transcription, BoltDB deduplication, CSV export.
- twitch-realtime-scraper: live tracking (viewers, title, category), HLS segment tracking, on-the-fly transcription.
- Multi-stream EventSub + IRC monitor with event processors (chat, follows, subs, engagement) and graceful shutdown.
Outcome
Working prototypes used to collect data; a personal research project that shaped how I write clean, concurrent Go.