Codebase RAG
Delivered a Workers + D1 retrieval pipeline over 1.2M LOC that serves accurate code answers in 420ms p95, boosting incident triage resolution by 37%.
Personal notes / Technical blog
I'm Roy Zhu. This is my personal site and technical blog, where I keep notes on architecture, developer experience, AI-assisted workflows, and the practical details behind building software.
Delivered a Workers + D1 retrieval pipeline over 1.2M LOC that serves accurate code answers in 420ms p95, boosting incident triage resolution by 37%.
Launched a Playwright + Workers co-browsing assistant that cuts enterprise onboarding time by 45% through scripted product walkthroughs.
Why Neo4j is overkill for a personal AI agent's memory, and how to implement fast, multi-hop context retrieval using a Breadth-First Search (BFS) in pure Python over SQLite.
Why pure LLM extraction fails at knowledge graph creation, and how a 3-tier cascade of exact aliases, embedding similarity, and string fallback keeps a local SQLite graph clean.
How membox chose its libraries, organized its modules, and used typing.Protocol to keep the core dependency-free while staying testable and extensible.
A local-first memory layer for AI agents: SQLite WAL for durable triples plus an in-memory graph for retrieval. No server, no vector store, no graph database.