Why we built it.
To prove a point about what production-grade AI infrastructure actually looks like — the unglamorous parts that separate a working AI system from a demo. Not a chatbot. Not a wrapper around an API. An operational pipeline with real failure modes, real observability, and a real human-in-the-loop interface.
Architecture.
Ingestion
- Python pipeline orchestrated by launchd with wall-clock-anchored scheduling (every 10 minutes, on the dot) for predictable cadence
- Disk-cached YouTube Data API v3 listings to stay under the free quota of 10,000 units per day, even with 16 channels polled continuously
- Per-video status state machine with TTL'd cooldowns to prevent retry storms after a YouTube IP block — failed fetches back off; the pipeline keeps running on other videos
- Configurable per-channel content windows (24 hours default, 7 days for high-signal channels) so the pipeline doesn't exhaust itself on content the user doesn't actually want
AI Processing
- Each transcript sent to Claude with a tuned system prompt that classifies relevance, extracts structured summaries, and rejects off-topic content automatically
- Skip logic: book reviews, music videos, channel intros, and other off-topic content are filtered before they consume tokens unnecessarily
- Output JSON-parsed and validated before being committed to disk state
Presentation
- Static HTML rendering for the read-only consumption view — fast, no JavaScript framework, indexable
- Interactive queue-management UI (in progress) for human-in-the-loop curation: drop, prioritize, mark not-news
What it demonstrates: We ship production AI infrastructure. Not chatbots, not demos — operational systems with proper error handling, rate limiting, scheduling, observability, and a clear separation between data collection, AI processing, and presentation. The same patterns apply directly to any client AI workload: document intake, customer-communication triage, compliance Q&A, sales-call summarization.