Case study · 2026

Personal project · Agentic RAG

startlight-brain-app

A "second brain" app with semantic search, pgvector embeddings, knowledge graph, and Web Clipper — built end-to-end as a solo project.

45 commits · semantic search pipeline with hybrid retrieval (cosine + top-K)

Role
Solo builder & architect
Stack
TypeScript · React Native / Expo · Supabase · pgvector · PostgreSQL · d3-force · OpenAI embeddings

Context

Personal knowledge management is broken: notes are siloed, search is lexical, and connections between ideas rely on manual tagging. The goal was to build a mobile-first “second brain” that surfaces semantic relationships automatically — think Obsidian meets semantic search.

Approach

Built with React Native (Expo) for cross-platform mobile, Supabase for backend, and pgvector for vector storage. The embedding pipeline runs on-device and syncs with a remote PostgreSQL/pgvector instance. A knowledge graph visualisation (d3-force) renders semantic proximity between notes. The Web Clipper lets users capture content from any browser; embeddings are generated and linked into the graph automatically.

Outcomes

  • End-to-end semantic search with pgvector: cosine similarity + top-K retrieval over personal notes.
  • Knowledge graph visualisation with force-directed layout showing semantic clusters.
  • Web Clipper with automatic embedding and graph linking.
  • Local + remote embedding sync for offline-capable operation.
  • 45 commits over 2 months — solo project from idea to shipped app.

Links