I design AI systems that agents can rely on:
MCP servers, open models on GPUs you control, and the cloud underneath.
Paris, France Β· AI architecture Β· Model Context Protocol Β· NVIDIA Β· Hugging Face Β· AWS Β· DevSecOps
|
I turn an API, a product or a workflow into tools any AI agent can call: typed inputs, guard rails that explain how to fix a bad call, stdio and HTTP transports, tests and docs. |
Open models such as Nemotron, Mistral and gpt-oss, served with vLLM or managed inference, from serverless GPUs to on-prem clusters linked by InfiniBand, and wired into the agents and editors your team already uses. |
|
Architectures that keep your data where you decide: European cloud, on-prem AI factories, RAG over your own documents, with a cost model for each option. |
AWS multi-account platforms, Terraform, Kubernetes, CI/CD, Security Hub and FinOps: the ground an AI system needs to run safely and at a known cost. |
An MCP server that lets AI agents build isometric Tiled maps: rooms, corridors, doors, traps, chests and spawns, validated and saved as files that open directly in Tiled.
- πΊοΈ 30 MCP tools, 3 one-call layout presets and 20 themes, 12 of them with bundled isometric sprites.
- π‘οΈ Built for agents: an invalid placement fails with a message that says how to fix it, and
validate_mapchecks the map before export. - π Works with any MCP client: Claude Code, Claude Desktop, Cursor, VS Code, Zed, OpenCode, and agents running NVIDIA Nemotron.
- π€ Live on Hugging Face: describe a map in any language and download it, or connect your agent to the hosted MCP endpoint. Paid model calls run on the visitor's own Hugging Face account.
claude mcp add --transport http tiled-map-builder https://4math-tiled-map-builder.hf.space/mcp \
--header "Authorization: Bearer hf_YOUR_TOKEN"
Dream cards for Teklovossen, the Mechanologist hero of Flesh and Blood, plus a small game simulator and a win-rate model, written as a step-by-step way to learn machine learning in Jupyter.
- π 38 fan-made cards along four tech paths (AI, Nano, Quantum, Bio), with concept art.
- βοΈ A readable game simulator: 800 random decks Γ 50 games each, against one fixed opponent.
- π Learning notebooks: sanity checks, feature engineering, the leakage trap, cross-validation, learning curves and a held-out test set.
- π An honest model: cross-validated RΒ² 0.67 and 0.68 on unseen decks, where the first version scored β1.33.
- π€ Next: an LLM through Hugging Face Inference Providers reads the card texts, and the fixed model goes to the Hub.
Fan-made and educational. Not affiliated with Legend Story Studios.
How to host generative AI and keep control of your data: research reports, reference architectures and Terraform templates, written in French.
- π 10 reports as standalone HTML pages, each with its sources and what is still to verify.
- πͺπΊ European cloud: Mistral on Scaleway in the Paris region, with inference modes, costs and GPU availability.
- π© On-prem AI factory: NVIDIA GPUs linked by InfiniBand, from Quantum-X800 to a local lab that starts on the ibsim simulator.
- ποΈ GPU clusters scheduled with Slurm, with a proof of concept run on a simulated fabric.
- 𧱠4 Terraform products deployable from GitHub Actions: a Mistral API gateway, RAG, agents and analytics, on a shared private, secure-by-default foundation.
| Project | What it is | Built with |
|---|---|---|
| nvidia-coding-agent-lab βΆ animated architecture |
An NVIDIA-specialised coding agent: Nemotron served with vLLM on serverless GPUs, wired into Zed and OpenCode, with NVIDIA's docs available over MCP. The same container runs on-prem, and a multi-node InfiniBand lab rehearses the cluster. | Nemotron Β· vLLM Β· Modal Β· MCP |
| aws-security-hub-soc | Security assessment across AWS accounts for SOC and audit teams: Security Hub findings plus Prowler audits. | AWS Β· Prowler Β· Shell |
| aws-finops-tools | A Docker-based toolkit for AWS cost management and optimisation. | AWS Β· Docker |
I'm moving my agent work into the open. The MCP servers and models behind my projects are being published here and on Hugging Face, one at a time, each with docs, tests and a live demo. β Star or follow to get them first.
| AI and agents |        |
| Languages |    |
| Cloud and platform |      |
| Delivery and security |    |
| AWS resources automated across multiple accounts |
applications shipped through CI/CD pipelines I built |
cloud infrastructure costs after optimisation |
AWS accounts covered by Security Hub |
Most of my professional work lives in private repositories; what you see here is the open part.
Building an AI agent, an MCP server for your product, or a private AI platform? I can help with architecture reviews, MCP server design and builds, and open-model deployments on cloud or on-prem GPUs.
"Between AI, DevOps and game design, I share my creations and take you behind the scenes of how they're made."