I spent six years building the invisible kind of software: a low-code platform and large-screen visualization systems at JD.com, then micro-frontends, a Salesforce rules engine and an AST-diff based E2E test selector at Compass. The pattern across all of it is the same — take a big opaque system and make it legible enough that other people can reason about it, and change it without fear.
In 2022 I went fully remote. Four years and two companies later I have a harder, less comfortable lesson: remote work replaces trust with process, and process starts measuring the wrong thing. I built a rules engine and a trading platform from zero, and twice lost merge rights — not over broken code, but because nobody in the room could see the work. I know now that I need a team in the same timezone and, honestly, in the same room.
So now I work on the part of AI infrastructure I can verify line by line: graphrag-ts, an open-source GraphRAG implementation in TypeScript on Postgres/pgvector — incremental indexing, Leiden communities, hybrid retrieval with evidence grounding, MCP server included. I write the internals down as long posts because that layer is under-documented. I am open to agent-engineering and frontend roles, and to a small number of consulting engagements on retrieval and knowledge-graph systems.