Projects

Pick a project to read about it. The notebooks and the election map are embedded below the summary.

California Political Knowledge Map

An LLM-built knowledge graph of California politics from news reporting

  • Built a pipeline turning CA political news into a Neo4j knowledge graph with source provenance
  • Made scrapers self-heal via JSON-LD validation and an LLM generator/critic selector loop
  • Filtered article relevance with keyword rules and an LLM agent guided by examples
  • Shipped a FastAPI Q&A app fusing full-text, vector, and LLM-generated, validated Cypher search
  • Iterated on prompts and code by reviewing failures and rerunning LLM evals, primarily on silver sets
  • Cypher
  • FastAPI
  • Logfire
  • Neo4j
  • Pydantic AI
  • python
Application data flow for the California Political Knowledge Map. Build the collection: collect and validate articles into SQLite; map the reporting with the Story Pipeline (coreference resolution, topic classification, Article Chunks, vector embeddings, entity and relationship extraction); store text and relationships in Neo4j with full-text and vector indexes. Find reporting for a question: ask a question; run full-text, vector, and relationship search in parallel, merge and rank, then assemble evidence by Source Article; read the result as a Source Guide written by the Chat Agent and saved in SQLite.

Live app: california-knowledge-map.akaisenghor.bottle.cloud