emanuele User

I’m an AI Product Engineer building production-grade LLM applications end to end—retrieval-augmented generation (RAG) pipelines, LLM evaluation harnesses, and full-stack product delivery. I focus on getting models to work in the real world: quality is measured, not assumed.

I shipped a Spanish regulatory RAG assistant with retrieval quality validated at Hit@8 0.98 / MRR 0.955 against a human-labeled golden set, plus a meta-evaluated LLM-as-judge pipeline. I built hybrid BM25 + dense retrieval, cross-encoder reranking, grounded answer generation with source citations, and a three-layer evaluation harness.

When evaluation logic broke, I diagnosed the root cause (true-negative rate 0.00 from threshold miscalibration and missing reference answers) and rebuilt it with reference-aware prompting and explicit fault enumeration. After revalidation, faithfulness TPR/TNR reached 1.00/1.00 and completeness TPR/TNR reached 1.00/0.875 on human-labeled data.

I bring hands-on rigor from a construction background and a habit of verifying work against real-world standards. I’m now seeking a remote product engineering role on a team shipping LLM features to users.