Staff Engineer at Seedtag, on Connected TV infrastructure. Show contextualization went from a blank page to 900M+ daily bid requests.
Still 70% execution, 30% strategy. Most of that execution runs through agents now.
Building since 2001. .NET first, then JavaScript and TypeScript, now mostly Python and Rust. Four domains along the way: e-commerce, IoT, AdTech, developer platforms.
In 2017 I was generating design systems from plain-language brand input, which was an odd thing to be doing at the time.
I've started things, too. Software companies, and before those a photography studio in Paris: 400 square metres, all natural light, four years.
One of them turned a games console balance board into a scientific measurement instrument, then ran it on a few hundred people across two cities. The software ended up assessing elderly people and children with disabilities.
The most useful one was never a company. Back when I was freelance in Paris I'd take jobs too big for one person, split them across a dozen others, and teach them the parts they hadn't done before. That's the closest thing to a Staff role I had before the title existed.
Most of that work lives in private repos and in client codebases I don't own. What's public here is the tooling I could pull out of it.
Self-healing pipelines. A corpus of millions of shows, continuously re-derived and kept current without anyone watching it.
A software factory. User interviews and product specs in, validated code deployments out, backed by dynamic knowledge graphs.
Both are agentic systems in production, which turns out to be a different problem from agentic systems in a demo.
Agentic tooling and MCP
- personal-shopper-agent - autonomous agent skill: category research, deterministic constraint auditing, primary-source verification, single-file HTML reports
- okapi-okf-knowledge-studio - TypeScript studio for Open Knowledge Format bundles. Explore the graph, audit and edit concepts, query it with an LLM
- mcp-twilio-sms, mcp-vapi-generic-caller, mcp-shopping-list - Model Context Protocol servers in Python
Documentation as code (old stuff, still used by some people)
- structurizr-gen-images and structurizr-pr-comment - GitHub Actions that regenerate C4 architecture diagrams on every pull request and comment the diff back onto the PR (demo)
- structurizr-cli-with-bonus - the Docker image behind them, with Git, Graphviz, jq and PlantUML
Developer tooling
- dead-code-finder - static analysis for unreachable Python
- inline-copy - reads files, formats them, drops them on your clipboard for pasting into a model that won't take uploads. Still useful!
Python, Rust, TypeScript, Kafka, Redis, Kubernetes, SQL. High-concurrency backends, event sourcing, distributed systems, agentic AI.
LinkedIn · Aix-en-Provence, France