A few systems worth writing up. Each one is a real problem, the design I chose, and what it changed.
RAG inference gateway
A gateway that fronts embedding and generation models with caching, batching, and per-tenant budgets so answers stay fast and costs stay honest.
Halved p99 latency under peak load.
Distributed payments ledger
A double-entry ledger that reconciles millions of daily transfers without drift. Event-sourced core, idempotent writes, and a replayable audit trail.
Cut reconciliation lag from hours to near real time.
Real-time analytics pipeline
A streaming pipeline turning raw events into query-ready rollups, with backpressure that protects the database instead of drowning it.
Sustained ingestion through 10x traffic spikes.