LLM application engineering
RAG, vector databases, agent tool-use, MCP servers, prompt and context engineering, lightweight evals. End-to-end systems, not toy demos.
15 years of production Python in studios where code had to ship under deadline pressure; now building the LLM application and agent infrastructure layer on top of it.
RAG, vector databases, agent tool-use, MCP servers, prompt and context engineering, lightweight evals. End-to-end systems, not toy demos.
Harnesses, custom CLIs, and skills around AI coding assistants so engineering teams can actually use them.
FastAPI backends, data pipelines, microservices. Python is the through-line; Go for single-binary CLIs, TypeScript where it earns its place, C++ at the engine layer.
Sounding-board for founders or early CTOs working on agentic workflows or LLM tooling. Low-volume, direct, practical.
Open Model Context Protocol server that lets AI agents act inside Autodesk Maya: read scene state, run tools, verify their own work.
Protocol-first, model-agnostic. Documented in a 3-part public case study including a real production-crash debugging walkthrough.
End-to-end RAG chatbot on Django and Celery.
Viewport UI framework for Autodesk Maya.
CLI that compiles Python codebases into Cython extension packages.
Go CLI that connects to your local bank to read balances and transactions from the terminal.
Python data pipeline on AWS that ingests 3D product models, processes them through Pixar USD, and stores queryable asset data in MongoDB.
Internal harnesses, custom CLIs, and bundled agent skills around Claude Code and OpenAI Codex.
Why Autodesk Maya needed its own Model Context Protocol server, and how the protocol-first design keeps it model-agnostic.
ReadHow agents call tools, verify their own work, and what the tool surface looks like in practice. With live transcripts.
ReadA real production-crash walkthrough where the agent identifies the offending node faster than a manual bisect.
ReadHow this writing surface is deployed. Cheap, fast, repeatable.
ReadI'm a senior Python engineer based in Belgium. My work centers on building LLM applications and the agent tool-use systems around them, owning production Python stacks, and rolling out agentic developer workflows for engineering teams.
Most recently I built an open-source MCP server that lets AI agents operate inside Autodesk Maya, with a 3-part public case study documenting it. I also built Kaiwa, a production RAG chatbot on LangChain and Pinecone with hybrid retrieval and Cohere reranking.
Before that, I spent 15 years shipping Python software in studios where code had to run reliably under deadline pressure: The LEGO Group, Scanline VFX, DNEG, and earlier studios in Belgium and the Netherlands.
Python is the through-line, but the toolbox is wider when it helps: Go for single-binary CLIs, TypeScript and SQL where they earn their place, C++ down at the engine layer.
Available for freelance engagements via my own company.
Freelance availability via my own company. Rates on request. Best by email; I read everything that lands.
bram.van.rompuy@gmail.com