Backend engineer. Python, FastAPI, Kubernetes.
- ML pipeline for commercial real estate price estimation. LightGBM, CatBoost and XGBoost ensemble at 15.3% MAPE, MLflow tracking on EKS, 300+ experiments. Promotion is decided by paired Wilcoxon and bootstrap CI, not a single split.
- An internal Slack assistant built on Claude API and MCP. Code analysis, natural language DB queries and Jira drafting, all in one channel.
- Retrieval for that assistant. bge-m3 embedding kNN routes intent and the LLM only breaks ties. I tested retrieval on about 100 real questions before adopting it, found grep-based agent search worked better for answering, and moved retrieval to routing instead of answer generation.
- Gold labels collected from real usage logs, used to calibrate routing confidence thresholds. Only validated examples get into the production index.
- One MCP server unifying Jira, Slack, Notion and Datadog tools, with a hand-written Markdown/ADF converter.
- FastAPI services on EKS with Helm, currently at 99.9% availability.
- MLflow: the tracking UI sat on a loading skeleton forever with no error, no console output and no network request.
MlflowRoutergates the entire app on one query settling, and aserver-inforequest that never responds had nothing to bound it, so it now times out into the fallback the code already had. #26027, merged - Same blank screen, different cause: React Query pauses every query when
navigator.onLineis false, and desktop Chrome can latch that process-wide while the network is fine. Bisected it across MLflow versions and browsers before touching any code. #25933, fix in #25961, merged - cmux (macOS terminal, Swift): made
setup.shfail early on a missing Metal or Rust toolchain instead of dying mid-build, and fixed the prerequisites it never documented. #12352 - Reported a native-fullscreen window bug there with a 4 Hz
CGWindowListprobe, turning "the screen looks broken" into exact frame coordinates and ruling out the pane layout. #11822
- Normalized tables with hundreds of millions of rows and tuned the queries. Reads got about 18x faster.
- Archived 900M rows of member statistics, cutting table storage by 90%.
- Moved EKS nodes to ARM Graviton and cut cost by 20%.