I build reliable backend systems and the AI services that run on them — and I've shipped merged code inside the LLM inference engine underneath.
Open to 2027 new grad SDE and AI Engineer roles in the US.
SGLang is a high-performance LLM serving framework. Across six merged PRs I worked its Apple Silicon / MLX backend — process lifecycle, scheduler state, profiling instrumentation, and the test infrastructure that keeps the backend honest.
| PR | What it does | Area |
|---|---|---|
| #30125 | Fix FakeOverlapScheduler test stub broken by forward_ct accounting |
Test infra |
| #29311 | Fix FutureMap relay unit test to use RelayPayload |
Test infra |
| #29217 | Fix step-bounded profiling for bench tools on Apple Silicon | Profiling |
| #28122 | Add Metal profiling hooks to the server profiler | Observability |
| #27190 | Emulate PDEATHSIG on macOS to prevent orphaned worker processes |
Process lifecycle |
| #26882 | Set canary_manager and materialize overlap-loop inputs on Apple Silicon |
Scheduler |
- expense-approval — Expense
reimbursement workflow with department-scoped RBAC, enforced in annotations and the
service layer. Live demo — sign in as
alice@acme.com/password123. - flashdeal — Java 21 flash-sale backend: Redis + Lua admission control ahead of the database, with Lua compensation when a RabbitMQ publish isn't acked.
- care-plan-rag — Retrieval-grounded clinical care plans: BM25 + kNN fused with RRF, and a two-layer eval gated in CI (recall@3 = 1.00, NLI contradiction ≈ 0.01).