01
Proof by adoption, not by claim
An open-source project with 300,000+ Docker pulls and around 1,400 stars. Real users, real bug reports, real backwards-compatibility constraints across 1,800+ source platforms.
AI Systems Engineer · Paris
I build AI systems that hold up in production.
Yann Orieult — at the intersection of generative AI, software architecture, distributed systems and infrastructure. Ten years across research, enterprise data platforms and production engineering, now spent building the systems that let AI do work rather than describe it. I design them, I ship them, and I run them.
- Docker image pulls
-
300k+
Docker image pulls
HomeTube, open source
- GitHub stars
-
~1,400
GitHub stars
from people I've never met
- production services
-
30+
production services
self-hosted, operated solo
- research → production
-
10 yrs
research → production
from physics labs to GenAI
What sets this profile apart
Building a system, not a demo
02
I build the tooling agents consume
Most GenAI profiles can call an API. I ship the other side of it: an MCP server, a REST API, a typed Python SDK and a CLI — designed, published and operated around one self-hosted engine, on one public contract.
03
I deploy and operate what I build
Thirty-plus production services on a self-hosted distributed infrastructure, with the observability, automation and incident recovery that implies — and nobody else on call. DevOps is not decoration around the product; it is what decides whether an AI system survives real users.
04
A quantitative foundation underneath
A double bachelor's in fundamental physics and theoretical mathematics, an MSc in AI from Sorbonne, and research carried through to publication in nuclear fusion simulation. Models are not black boxes to me.
Selected work
Things that exist and run
- HomeTube Open-source media service shipped as a Docker image, adopted at a scale that imposes real product constraints.
- Content A self-hosted engine exposed through an MCP server, a REST API, an SDK, a CLI, a browser extension and three web UIs — one core, one contract.
- Agent-to-corpus gateway An SSH forced-command gateway letting remote agents query a private corpus in natural language — no shell, no path disclosure, no file transfer.
What comes next
A bigger problem, not a safer version of this one
I am looking for an environment where I can broaden my impact without stepping away from the engineering reality — real users, demanding production constraints, and end-to-end responsibility. The kind of role where the difficulty is part of the interest: scale, critical systems, emerging subjects, and owning the outcome.
Based in Paris, open to remote and hybrid, with a preference for French-speaking Switzerland when there is time on site. Available immediately, no notice period.