Martin Cousseau, AI Engineer.
I build LLM systems and the evaluations that tell you whether they are ready to ship.

§01
What I deliver
- #1 · STRATEGY
- #2 · DATASETS
- #3 · HARNESSES
- #4 · HUMAN
- #5 · AGENTS
- #6 · CONTINUOUS
- #7 · RISK
- #8 · THE METHODmRead how I work
// not sure where to start?
The method behind the seven guides.
§02
Flagship writing
Semantic Retention and Extreme Compression in LLMs: Can We Have Both?
arXiv 2505.07289 (opens in a new tab) · IJCNN 2025, Rome
Companion note
What Survives Compression
“Ask what meaning survived, not how fluent the remainder sounds.”
Extraction Arena
A strong average can hide a procedure the model gets entirely wrong. Gate the field you fear, not the mean.
// note + walkthrough + open-source code
† 2025 International Joint Conference on Neural Networks (IEEE), Rome.
§03
Stack
- Build
- Python (pipelines & evals)
- React (interfaces)
- Next.js (web apps)
- LLM ops
- Langfuse (tracing & evals)
- Models
- Anthropic (Claude models)
- OpenAI (GPT models)
- Gemini (Google models)
- Grok (Grok models)
- DeepSeek (open weights)
- GLM (Z.ai open weights)
- Kimi (Moonshot AI models)
- Inference & cloud
- Fireworks AI (open-model inference)
- Microsoft Azure (enterprise cloud)
- Microsoft Foundry (managed models, formerly Azure AI Foundry)
- Ship & data
- Docker (containers)
- Vercel (deploys)
- Supabase (backend)
- PostgreSQL (database)
- Claude Code (coding agent)
- GitHub Copilot (pair programmer)
- Pi (terminal coding agent)
- oh-my-pi (Pi fork)
- Hermes Agent (agent harness)
- herdr (agent multiplexer)
- Omarchy// my main OS
§04
About
// the short version
French AI Engineer in Warsaw. I build LLM systems and the evaluations that show whether they are ready to ship, and I co-authored an IJCNN 2025 paper on semantic retention under LLM compression.
- base
- Warsaw
- speaks
- English / French
- paper
- IJCNN 2025 · arXiv 2505.07289
- open to
- B2B engagements
§05
Contact
// one line, one action
Tell me what you’re about to ship.
$ mail contact@martincousseau.com