Deep Cuts

AI-powered media intelligence pipeline. Ingests YouTube shows and podcasts, transcribes them, runs two-pass LLM analysis, and surfaces cross-episode connections through semantic search and auto-generated infographic cards.

Tech Stack

Worker: TypeScript, Anthropic SDK (Claude Sonnet), OpenAI Embeddings, PostgreSQL + pgvector, Zod

Frontend: Next.js 15, React 19, Tailwind CSS, D3, Satori (JSX to SVG)

Features

• Two-pass LLM pipeline: Pass 1 extracts structured data per episode (TLDR, entities, quotes, sentiment, keywords). Pass 2 connects episodes into cross-show thought threads and thematic bridges.
• Semantic search: pgvector HNSW indexing on chunked transcript embeddings for fast cosine similarity retrieval, no external vector database needed
• Template-based infographics: 7 card templates (stat-card, quote-highlight, topic-web, timeline, comparison, data-grid, flow) rendered via Satori for pixel-perfect output
• Smart ingestion: YouTube captions via free API (Whisper fallback), podcast RSS parsing, thumbnail color extraction for visual theming
• Schema-validated outputs: every LLM response validated against Zod schemas before database insertion

Architecture

YouTube/Podcasts → Ingest → Transcribe → Analyze (Pass 1) → Embed → Connect (Pass 2)
                                                                        ↓
                                              Frontend ← PostgreSQL + pgvector

Development

docker compose up -d   # PostgreSQL + pgvector
npm run db:migrate     # Apply schema
npm run dev:worker     # Start worker
npm run dev:frontend   # http://localhost:3002

---

Part of the travisBREAKS portfolio.