I work on AI platforms: agent runtimes, LLM gateways, authorization, sandboxed execution, and the Kubernetes and on-prem infrastructure they run on. I learn the most from the places where application code meets framework internals — async execution, queues, memory behavior, connection pooling — and from production failures I didn't see coming.
Some of the problems I've been working through on this project:
- Relationship-based access control. Domain facts stay in PostgreSQL and are projected into a ReBAC graph by a background worker. Checks fail closed until a membership change has been applied, so users read their own writes without every check paying for full consistency.
- Recording every change. Mutations run as typed transactions or recoverable sagas, with idempotency keys, before/after snapshots, an append-only audit log, and revert plans that stop when the current state has drifted.
- Letting tooling catch mistakes. Repository checks cover per-handler authorization policy, database access ownership, bounded request I/O, and migration safety. I try to treat a fix as done only after the affected path has actually run.
At MaiAgent I worked on the migration from LlamaIndex to LangChain/LangGraph, an OpenAI-compatible LLM gateway, and a fair amount of production debugging across CPython, aiohttp, Django ASGI, PgBouncer, and Celery — with a lot of help from a great team.
LlamaIndex PRs · Lightpanda #4080 · OpenTelemetry JS #6215 · AI Elements #125 · NeoForge #3260
Private monorepo counters updated 2026-11-04 (UTC). Counts cover authored merged PRs and all issues since repository creation, workspace package/app manifests, numbered ADRs, and paired bilingual documents under docs/ (excluding README indexes).
Public GitHub counters updated 2026-11-04 (UTC).