I fix correctness bugs in ML infrastructure: inference servers, training frameworks, and benchmark harnesses. I study computer science at NYU and am working toward a mathematics minor (declaration in progress). I am looking for 2027 new-grad roles in ML infrastructure and backend engineering.
Counts as of October 2026. Many of these changes were developed with AI coding assistants.
- llm-d (benchmark, 29 merged; inference simulator, 11): report and metric fixes, e.g. keeping each endpoint's request count in the EPP request distribution.
- EvalScope (55 merged): benchmark prompt and scoring fixes, e.g. including the passage and candidates in ArxivRollBench cloze prompts.
- Xinference (34 merged): model-catalog fixes, e.g. restoring the official Qwen2.5-Coder chat template.
- TorchTitan (9 merged): training correctness, e.g. using Llama 3.2's RoPE scaling factor for the 1B and 3B flavors.
- GuideLLM #1194: fixed percentiles landing one rank high from cumulative-sum rounding.
- Hugging Face Transformers #47238: fixed half-precision
torch.compilefailures in DETR-family position embeddings and added a compiled-dtype regression test. - Home Assistant (31 merged): integration startup and error-handling fixes, e.g. Tuya startup recovery.
- AgentCI Guard: an experimental TypeScript static analyzer for GitHub Actions workflows that run AI agents. It models workflow permissions and reachability and emits diagnostics with explicit analysis limits. Release v0.7.0 · Demo.
- TraceBench: a replay-sufficiency benchmark for ML decision evidence, with reproduction instructions and scoped experimental evidence in the repository.
- CloudTune: an early ML infrastructure project for supervised fine-tuning on customer-operated infrastructure, failure recovery, and portable signed execution receipts. The public research preview includes an illustrative receipt-verification demo; the application and source remain private.
How should an inference change be evaluated when throughput, latency, output requirements, and run-to-run variation all matter? I am interested in building reproducible experiments and failure cases that make those trade-offs explicit.