Now showingCH 01
Throughline, Local-First AI Second Brain
- Built a full-stack local-first app, a FastAPI backend over a hybrid sqlite-vec (vector) + FTS5 (keyword) index with transactional writes and SSE streaming, paired with a crafted React 19 + Tailwind multi-screen UI (onboarding flow, screens, custom typography).
- Ran three local Ollama models, text reasoning, vision (multimodal image reads), and embeddings, behind an LLM-ops manager that keeps a single heavy model resident within a fixed ~16 GB RAM budget.
- Engineered a pausable, read-once intake worker whose idempotency is enforced by content-hash upserts and verified by dedicated idempotency and determinism test suites.
- Guaranteed deterministic, schema-validated (Pydantic) generation cached by source signature, temperature, and seed.