Systems I've Built

I wrote the Raft consensus protocol from scratch in Go, then built a key-value store on top of it. Five nodes elect a leader, replicate a log, and recover cleanly from crashes. I threw simulated network partitions at it with over a million entries loaded and it kept strong consistency the whole time. Storage runs on RocksDB with MVCC versioning, writes clock in around 10k per second, and Prometheus tracks throughput and p99 latency. Each node ships in its own Docker container.

A job queue in Go that spreads work across five worker nodes and chews through about 10k tasks an hour. Jobs carry priorities, a token bucket keeps anyone from flooding the queue, and work that keeps failing lands in a dead-letter queue instead of silently disappearing. Prometheus and Grafana watch queue depth, latency, and failure rates, and I wrote runbooks for the alerts that actually matter, like workers getting stuck.

Search that understands what code does instead of just matching names. It blends BM25 keyword retrieval with embeddings stored in Postgres through pgvector, which cut irrelevant top results by 40% compared to a vector-only baseline on a query set I labeled by hand. There is an offline harness that scores recall@k and MRR so I can tell if a change actually helped, and a Redis cache that keeps p95 latency 35% lower under load. It indexes a 500k-line codebase.