AI engineer focused on LLM evaluation, RAG systems, and fine-tuning. I find bugs by running tools against real workloads, then fix them upstream — three merged into libraries that ship in production.
| Project | Contribution |
|---|---|
| Haystack | ✅ Merged — #12726 fix: strip UTF-8 BOM in text, CSV, Markdown and JSON converters. A BOM-prefixed file silently carried a zero-width U+FEFF into the first Document. Reviewed across five rounds by a deepset maintainer. |
| Arize Phoenix | ✅ Merged — #14995 fix(traces): render ReasoningPart in gen_ai message flattening |
| pydantic-ai | ✅ Merged — #7290, written from my issue #7281 and verified on Windows at the maintainer's request |
| Braintrust autoevals | PR #209 — AnswerRelevancy silently ignores embedding_model · Issue #211 — cross-model embedding-cache poisoning. Both found by running autoevals against my own RAG eval harness. |
Citation-backed RAG chatbot over a 552-document biomedical corpus. Multi-stage retrieval (HyDE → two-pass Pinecone → reranking, with a cross-encoder fine-tuned on hard-negative-mined pairs), tiered data-cleaning pipeline, LLM-as-judge eval harness scoring 9.64/10 answer quality, ~$0.008/query in production.
metal-llm — LLM-generated guitar tablature
Fine-tuning a small LLM to write heavy-metal guitar tab (DadaGP token format) — QLoRA on 8 GB VRAM. Built a before/after eval suite under one frozen config (validity, decodability, novelty): tab validity 2.1% → 99.5% over the base model. The training split is cluster-based because the official one leaked — 578 duplicate clusters straddled train and validation, which would have faked the result. Then I audited my own headline number: feeding the model a genre label it had never been trained on produced the same output as a trained one, proving the conditioning metric measured nothing — written up in the README. Pilot adapter on Hugging Face. Includes an audio demo where the model continues a real intro into an original solo.
📫 Open to early-career AI/ML engineering roles — evals, RAG, LLM tooling.