Personal Projects (Side)

Designed a personal workflow framework that automatically manages AI coding agents' context, cost, rules, RAG, and browser runtime, deployed across 3 harnesses

• Growing markdown knowledge base made manual context discovery by agents inefficient — Adopted a local-embedding (BGE-M3) vector RAG over cloud vector DBs (e.g., Pinecone) — chose local to auto-reindex immediately on markdown changes and avoid sync latency/cost; combined sqlite-vec with FTS5 (RRF fusion) for accuracy, exposed via MCP for direct agent search, Indexed 548 files / 3,347 chunks, 0.7s warm query, all 5 gold test queries hit within top-2/3
• RAG was an opt-in reference tool, so evidence search was missed during investigation and review prompts — Moved to a default UserPromptSubmit trigger and ported it across 3 harnesses so prompts of 12+ characters call search_harness automatically, Passed 15/15 default-on tests across 3 harnesses and hit the target on the first search in a real-prompt smoke test
• Resume and portfolio JSON were outside the RAG search scope, leaving JD customization without evidence search over those surfaces — Implemented portfolio/resume JSON chunkers, target=resume indexing, path-based auto-reindexing, and federated search, Indexed 109 resume-corpus chunks, passed 32 tests, and verified federated search across 4 DBs
• 95% of AI agent token cost came from hidden areas (system prompt, subagents), making it untrackable — Built a cost analyzer that decomposes session transcripts into 8 categories instead of simple session/model aggregation, which could not trace hidden costs — added residual tracking to consolidate multi-machine, multi-tool usage into a single profile, Made 95% hidden cost visible by category; found and fixed a multi-instance usage-aggregation bug
• Codex Browser and node_repl runtime trust boundaries and fallback order were unclear, making browser automation unstable across sessions — Documented node_repl MCP trust boundaries and environment-native Browser/Codex/Chrome fallback order, then reflected Chrome plugin activation in the launcher generator, Registered node_repl across kh/gp/gd harnesses and established browser runtime trust-boundary plus fallback policy

Built the deployment and scheduling infrastructure for a side project that auto-collects and normalizes job postings

• Free-tier hosting lacked cron support and the managed DB's free tier auto-paused, blocking scheduled jobs — Adopted a self-hosted Mac mini workflow-automation (n8n) scheduler instead of hosting cron, and standardized warm-up endpoints and secrets, Kept the free tier alive via daily DB warm-up and restructured manifest-based sync