Traces

At ChessMood I built the backend for game analysis. I designed the ingestion pipeline: an outbox, ordered queues, and idempotent SQS workers. I also built Lab, a shared store for games, engine evaluations, and metadata.

As principal maintainer of the Go chess library, I rewrote more than half the codebase. PGN parsing dropped from 26 seconds to 0.888 seconds in multithreaded mode; replacing the legacy Python parser gave a further 14× speed-up.

I processed 300+ hours of chess video. OpenCV and FFmpeg pre-filters made the board-recognition model roughly 10× faster, and made frame-accurate position search possible.

At Basalt, I redesigned a trace platform processing roughly one terabyte of telemetry per day. I moved storage toward Kafka, ClickHouse, and Quickwit, then reshaped ClickHouse queries: work that took dozens of seconds and hundreds of gigabytes of RAM now completes in under a second.

I built a custom OpenTelemetry Collector extension that adds authenticated, transformed ingestion. I also fixed multi-day traces that broke the original partitioning assumptions.

I led an internal workshop on refactoring toward TDD, clearer domain boundaries, CQRS where it helps, and end-to-end coverage across Go and TypeScript services.