Judgment over generation.

I build productionAI systemsfor domains whereAI systems for domainsa wrong outputwhere a wrong outputis expensive.

Grenoble, France · Minneapolis, Minnesota · Models

Models are cheap. Taste, evals, and architecture are not.

01 / Selected systems

The work is in the constraints.

Different domains. Same standard. If being wrong is expensive, the system has to prove itself.

01Production LLM platform / Nophin

Problem: A production workflow needed model outputs to be consistently correct, not merely plausible.
Constraint: The platform had to remain cloud-agnostic and move at startup speed without outsourcing reliability to one provider.
Decision: Built around explicit evaluation sets, targeted fine-tuning, and portable inference paths using Ray Serve and Modal.
Result: 100% accuracy on the defined evaluation set, with a repeatable path from experiment to production inference.

02Mission-critical industrial control / Phaidra

Problem: Industrial operators needed continuous optimization without compromising the systems already running the facility.
Constraint: The data layer had to absorb 5,000 points per second, roughly 30 GB per day, behind a 99.9% uptime target.
Decision: Designed real-time caching for model training, deployed reinforcement learning on ML accelerators, and instrumented the path with Grafana and Prometheus.
Result: 20% higher operating efficiency while maintaining the 99.9% service-level objective.

03Applied RL research / AWS-funded ADAS-E

Problem: Emergency vehicles need to reach incidents faster while still reasoning about traffic rules and environmental safety.
Constraint: The policy had to optimize competing objectives in CARLA and learn from expensive simulation runs across limited GPU capacity.
Decision: Combined bird's-eye sensor data, custom reward functions, tuned DQN variants, and distributed training across five GPUs.
Result: 90% better emergency-response performance and human-level driving performance; the work earned AWS research funding.

02 / Working with AI

I use AI to run more of the stack. I do not hand it the judgment. Tools change the speed of production. They do not change who is accountable.

01Cover more of the operation: search the problem, pressure-test assumptions, and reach a first prototype faster.
02Generate implementations, tests, migrations, and failure cases — then review them against the system.
03Automate the mechanical work so the remaining time goes to taste, evals, and architecture.

01Outsource the standard. If I cannot say what good looks like, the model cannot either.
02Ship unreviewed output into a high-consequence path, or treat fluency as evidence.
03Replace ordinary code with an agent when deterministic software is clearer, cheaper, and easier to test.

AI changes the cost of producing. It does not change the cost of being wrong.

03 / Notes

Working notes

Short notes on model behavior, eval design, agents, and architecture — published while the thinking is still being tested.

Long form

Longer write-ups on reinforcement learning, language-model reasoning, computer vision, and the decisions that produced the numbers.