Python/backend engineering · inference tooling · evaluation reliability
I build tools that make model-serving runs and product answers inspectable: what was requested, what happened, what evidence supports the result, and where the claim stops. Portfolio · LinkedIn
- #1434 — Preserve terminal benchmark failures: fixed missing-record parsing and debug logging so failures reach the controller.
- #1448 — Count completed requests from phase records: corrected a case that reported 74 completions for 3 requests; added regression coverage.
- #1447 — Group MLflow diagnostics under
system/: separated system diagnostics from inference metrics with tests and migration guidance.
Review the merged PRs upstream. As of 9 October 2026, #1463 and #1464 are open PRs under review.
- Inferdrome — Python pipeline for repeatable vLLM/SGLang serving experiments and offline-checkable evidence. Its A100 baseline is a bounded single-GPU result, not a capacity claim.
- Ablatrix — evidence-grounded product answers and a reviewer-driven revision loop. The pinned paid batch records 20/20 completed calls and 52/52 exact citation-quote matches; no answer-quality gain has been established.
- ExitSpec — Python acceptance workbench that freezes a criterion, checks evidence, and returns a scoped verdict. Its guided example uses synthetic cases, not a customer benchmark.