I build full-stack apps with Python and FastAPI on the backend and React and Next.js on the frontend, and I specialize in AI agents that stay under control. In Compass the model only extracts, and plain Python makes the call. In Nexus, nothing reaches a pull request until a person approves it.
Looking for my first developer role · Spain (on-site or hybrid) or remote.
Compass reads the daily notices from PLACSP, Spain's public procurement platform, keeps the few that fit a company and tells it whether it can bid, quoting the clause and page behind every requirement.
- The funnel never calls a model. Of the final 10 matches in the dev corpus, 6 came only from the vector retriever: they share meaning with the company profile but not its words.
- Every quote is looked up in Python on the page it cites, and the analysis shows how many passed.
- CI runs ruff, mypy in strict mode and every test except the five that need the real PLACSP corpus, against real Postgres and Redis.
Python · FastAPI · PostgreSQL · pgvector · SQLAlchemy · Celery · Redis · LangGraph · Langfuse · TypeScript · Next.js · Docker
Nexus reviews code with specialist agents. You point it at a GitHub repo or paste code, choose what to check, and a LangGraph router runs only those specialists, in parallel. A synthesizer turns their findings into one report.
- My own MCP server runs as a separate service and gives the graph three GitHub tools.
- Posting to a pull request pauses the graph until you approve it.
- If a specialist fails, for example on Groq's rate limit, the rest keep going and the report is built from what they found.
- The dashboard shows each agent's progress live, over WebSockets.
Python · FastAPI · LangGraph · MCP · PostgreSQL · Redis · WebSockets · TypeScript · Next.js · Vitest · Docker