emanuelecz
Remote: Yes, remote only
Willing to relocate: No
Technologies: Python, FastAPI, Celery, Postgres/pgvector, SQLAlchemy, Alembic, Redis, LangGraph, MCP, LangChain, Chroma, RAG (hybrid BM25 + dense, cross-encoder reranking), LLM-as-judge evals, Docker, GitHub Actions
Résumé/CV: https://www.emanuelecz.dev/emanuele-czofei-cv.pdf
Email: [email protected]
Self-taught AI engineer, two years building, four months at a startup. Strongest at the unglamorous half: retrieval that retrieves, evals you can trust, knowing how a system fails.
Built a Spanish occupational-safety RAG pipeline over the INSHT/NTP construction-safety corpus — hybrid BM25 + dense retrieval with a Voyage cross-encoder reranker, and an eval harness over a 50-question golden set with reference answers. Retrieval is the easy half: Hit@5 1.00, MRR 0.94, Precision@5 0.72.
The part I'd actually talk about is the judge. Rather than quote its averages (faithfulness 4.86/5, and so what), I hand-labelled 15 of the 50 answers good/bad and scored the judge against my labels: TPR 0.82, TNR 0.25. It approves three of the four answers I rejected. So the retrieval numbers mean something and the judge numbers don't yet — it's too lenient at the threshold I picked, and I haven't fixed it. Measured, not solved, and I'd rather say that than quote the 4.86. A threshold gate on the metrics is written but not merged. Small corpus, ~20 documents, and I say so. https://github.com/Emanuelo06/construction_rag
Building in public: Docent, issue triage for pydantic/pydantic. ~1.5k resolved issues ingested as a hybrid RAG corpus via GitHub GraphQL fetchers and Celery pollers, hybrid keyword + vector retrieval with reranking and score merging, three of five agent phases live (intake → research → draft), research calling the GitHub MCP server as a client. The judge gate, escalation, cost metering and injection defence are designed and not yet built. https://github.com/emanuelecz/Docent
Looking for a small team doing agents, devtools, or AI support tooling. EU citizen, Spain-based, EOR or contractor.
Remote: Yes, remote only
Willing to relocate: No
Technologies: Python, FastAPI, LangGraph, MCP, RAG, Postgres/pgvector, Redis, Celery, Docker, LLM evals, TypeScript, Next.js
Résumé/CV: https://www.emanuelecz.dev/emanuele-czofei-cv.pdf
Email: [email protected]
Self-taught AI engineer, two years building, four months at a startup. Strongest at the unglamorous half: retrieval that retrieves and evals you can trust.
Shipped a Spanish occupational-safety RAG assistant, Hit@8 0.98 / MRR 0.955 on a human-labeled golden set, evals gating CI. Its LLM-as-judge was passing everything at TNR 0.00, so the suite was decorative, until I traced and rebuilt it to TPR/TNR 1.00.
Building Docent in public: an autonomous support engineer for pydantic/pydantic. ~2k closed issues in pgvector, five-phase agent, LLM-judge gate, eval splits gating CI, cost metering, prompt-injection defense, MCP client and server. https://github.com/emanuelecz/Docent
EU citizen, Spain-based, EOR or contractor. Looking for a small team doing agents, devtools, or AI support tooling.
Remote: Yes, remote only. EU contract, or US via EOR (Deel / Remote.com). Registered freelancer, can also invoice B2B.
Willing to relocate: No
Technologies: Python, LangGraph/LangChain, RAG (hybrid BM25 + dense retrieval, VoyageAI embeddings + reranking, ChromaDB, pgvector), LLM evals (golden sets, LLM-as-judge, judge meta-evaluation), Claude + OpenAI APIs, FastAPI, PostgreSQL, TypeScript/React/Next.js, Docker, GitHub Actions
Résumé/CV: https://www.emanuelecz.dev/emanuele-czofei-cv.pdf
Email: [email protected]
AI Product Engineer focused on production LLM systems with measured quality, not vibes. Recent work:
- Spanish occupational-safety RAG assistant over the INSHT corpus: hybrid retrieval (BM25 + voyage-3 + rerank-2.5), Hit@8 0.98 and MRR 0.955 against a 50-item human-labeled golden set. Rebuilt the LLM-as-judge layer from TNR 0.00 to TPR/TNR 1.00/1.00 on faithfulness by fixing threshold calibration and adding reference context, with meta-evaluation gating on top. Evals run in CI and fail the build on regressions.
- RepoDoctor: LangGraph agent that clones a repo, auto-detects the stack, runs its test and lint suites in sandboxed subprocesses, and returns a structured diagnosis grounded in real command output. Streaming Next.js frontend.
Previously shipped auth flows in Next.js/TypeScript at a Swedish startup. Looking for a full-time role shipping LLM features end to end: retrieval, agents, evals, and the product surface around them. Live demo and case study: https://www.emanuelecz.dev