Backend systems and
ML infrastructure, built from scratch.

A price-time-priority limit order book on a lock-free single-writer pipeline: LMAX Disruptor ring-buffer intake, cache-line-padded sequences, object pooling, and off-heap serialization via Panama MemorySegment. Every headline number is measured under a realistic mixed order flow (multi-level sweeps, partial fills, cancellations), not a synthetic always-crossing best case, with the cheap-workload figures reported separately rather than quietly headlined.

A distributed, partitioned, replicated commit log — hand-built from the disk up. Segmented append-only log with sparse index and CRC, ISR replication, epoch-based leader election, idempotent producer, and group-commit batching. Zero external dependencies on the data path.

A benchmark-proven retrieval-augmented generation system over a 6M+ vector Wikipedia corpus. Hand-profiled with py-spy — the flamegraph exposed a batching bottleneck, yielding a 3× QPS gain. Agentic multi-hop reasoning (iterative retrieve-reason-retrieve with self-reflection) lifts HotpotQA exact-match by +15 points over single-shot RAG.

A real-time multi-agent architecture review system. Eight specialised LLM agents — planner, four parallel specialists (scalability, security, cost, maintainability), conflict detector, critic, synthesizer — run concurrently via LangGraph fan-out and stream verdicts over WebSocket. A human-in-the-loop revision loop re-runs the synthesizer without re-invoking the critic.