the drifting marks behind this text are a flocking simulation — the same three rules the swarms at work run on.
about
Hello! I'm Erwin. I architect robotics and AI systems — which in practice means I spend my time on the seams: the places where a planner meets a controller, where a model meets a tool, where a fleet meets the operator who is supposed to trust it. My path started in Nantes, France with electronics and embedded systems, detoured through backend engineering, and landed where I always wanted to be: teaching fleets of robots to make decisions together.
Today I lead autonomy teams at SIRB.AI and the Technology Innovation Institute, where nine engineers and I build decentralized autonomy frameworks for heterogeneous drone swarms — navigation, perception, orchestration, and the infrastructure that keeps it all flying. On the side, I build agentic AI and blockchain tooling at Unchained Labs.
The thread through all of it is the same instinct: push the decision to the edge, make the process reproducible, then get out of the way. It's why my drones have no leader, why my agents get tools instead of instructions, and why my laptop rebuilds itself from a text file. When I'm not shipping autonomy stacks I'm playing basketball, watching films, or chasing ranked ladders.
Here are a few technologies I've been working with recently:
// portrait pending — swarm standing in
experience
Eight years, four countries, and — right now — three roles running at once. Select a row for the detail.
2026 — Present · Abu Dhabi, UAE
2024 — Present · Remote
2022 — Present · Abu Dhabi, UAE
2021 — 2022 · Odense, Denmark
2020 — 2021 · Nantes, France
2019 · Coimbra, Portugal
2018 — 2021 · Bordeaux, France & Remote
projects
Three altitudes, one idea. A swarm decides without a leader; an agent decides with tools instead of a human; a machine rebuilds itself from a text file. Different problems — the same instinct about where decisions belong.
in the field
Decentralized, acceleration-based flocking for drone swarms, inspired by how starlings murmurate. Each agent decides from local observations only — no leader, no central brain — yet the flock moves as one. Published at IEEE/RSJ IROS 2024, and the reason the hero of this site is a boids simulation.
in the loop
A Model Context Protocol server that gives AI assistants read-only access to a Trading 212 investing account — balances, positions, dividends, pies, and the instrument catalogue. Strictly no trading: the HTTP client has no code path that issues anything but a GET, which is a design constraint rather than a policy note.
on the metal
One curl command turns a bare Ubuntu box into my entire working environment — shell, toolchains, containers, editors, and dotfiles via chezmoi — then hands off to the app layer. Ansible underneath for idempotency, a typed Python CLI so every generated command is auditable before it runs, and a small Flask control plane for composing a config without hand-editing YAML. It dry-runs with --check, and it is how every machine I touch comes to exist.
Serves local models on whatever hardware is actually in the box — vLLM with tensor parallelism on a GPU rig, Ollama with GGUF quantization on a Jetson or a bare CPU — behind one OpenAI-compatible URL. Provisioning detects the tier and steps back; the harness owns engine choice.
The other half of setup: the apps, ports and compose stacks, plus the dotfiles source. One boundary, strictly held — if it installs a toolchain it lives in setup, if it adds an app it lives here.
Agentic tooling that keeps documentation alive: watches a codebase, detects drift, and opens documentation pull requests on its own.
Agentic orchestration platform — coordinate fleets of AI agents the way you'd orchestrate services.
An MCP server over a public-records API, so an assistant can query the archive — persons, documents, flight logs — directly instead of being told about it.
Real-time jitter measurements under ROS 2 on the inverted pendulum case — tooling and a published preprint. Xenomai/XDDP underneath, because soft real-time isn't real-time.
A survey of the multi-agent pathfinding landscape: taxonomy, solvers, and the open problems — the reading that shaped how I think about coordinating fleets.
2D/3D pathfinding, a cascaded controller, and an extended Kalman filter for autonomous flying vehicles.
A C++17 pathfinding library — A*, Dijkstra, DFS/BFS and friends — with a clean interface for grid worlds.
contact
I'm not actively looking, but the inbox stays open — a question about swarm autonomy, an open-source idea, or just to say hi. I'll get back to you.