ML infrastructure engineer. I build the data and ML platforms that models train and serve on, and fix correctness bugs in the open-source LLM stack.
- Now: search signal infrastructure at Walmart Global Tech: real-time Flink and Kafka pipelines on GCP, a shared feature platform for ranking and personalization, and drift monitoring across 100+ production signals.
- Going deep on: LLM inference and agent infrastructure: KV-cache capacity, batching, and long-horizon agent evaluation.
28 merged pull requests across 13 projects, mostly shared-state, resource-leak and serialization bugs.
| Area | Merged |
|---|---|
| LLM pipelines | Haystack (6) |
| Fine-tuning and LLM gateways | PEFT (3), LiteLLM (2) |
| Agent and LLM evaluation | Inspect Evals (3), Harbor (3), HELMET, inspect-evals-lint |
| ML platform and data | MLflow (2), Feast, Airflow, Quix Streams |
| Forecasting | NeuralForecast (2), StatsForecast (2) |
A few I'm glad landed:
- PEFT #3774: LoRA
merge()on MultiheadAttention applied theout_projdelta twice. - LiteLLM #30271: aiohttp connections leaked when a stream ended abnormally.
- Haystack #11931: deserialization mutated the caller's data dict.
In review: vLLM, TRL, LlamaIndex, LiteLLM, lm-evaluation-harness, lighteval
- inspect-brokenmath: Inspect AI implementation of BrokenMath (NeurIPS 2025), measuring sycophancy in LLM theorem proving across 451 adversarially falsified olympiad problems.