🪷 Kosha

Fully Open Source Personalized Content Streamer

---

🚀 Minimal MVP Roadmap (4–6 weeks)

Week 0: Repo + Infra

• Scaffold monorepo (backend, frontend, connectors)
• Docker Compose environment
• Postgres + Redis + MinIO

Week 1–2: Ingestion + Storage

• Implement 3 connectors (RSS, YouTube, newsletters via Mailgun webhook or IMAP parsing)
• Parse → clean → store core schema + raw content
• Add transcript extraction for YouTube/podcasts

Week 2–3: Embeddings + Basic UI

• Hook in embedding model, store vectors in pgvector
• Build simple React UI showing daily digest from last 24h
• Add Obsidian export endpoint

Week 3–4: Feedback & Ranking

• Add explicit feedback API (save, highlight + reason)
• Basic ranking: recency + source weight + embedding similarity

Week 4+: Iterate

• Add Celery tasks
• Scheduled digest generation
• Analytics dashboard
• Begin gathering feedback for Phase 1 ML

---

⚠️ Danger Zones & Mitigations

• RLHF too early → wait for high-signal feedback (≥ 500 explicit examples)
• Over-indexing content → prune old vectors, compress, or store reduced embeddings
• Connector maintenance → build adapters + monitoring/tests per connector

---

✅ Quick Checklist to Start Coding (Actionable)

• Initialize repo + Docker Compose (Postgres, Redis, MinIO)
• Define DB schema (users, sources, items, interactions, feedback)
• Implement RSS & YouTube fetchers + HTML → Markdown parser
• Add sentence-transformers embedding job and pgvector integration
• Build FastAPI endpoints for digest and feedback
• Minimal React UI to view digest, mark likes/highlights, export to Obsidian

---

Full Architecture Details

Workflow