Senior AI Consultant at Wavestone · LinkedIn
🇫🇷 Version française en bas de page.
I build AI tools that run on ordinary machines, measure what they are really worth, and make visible how they work.
- Local, frugal AI: small language models (SLMs) on a regular PC, with no GPU and no cloud dependency, including carbon footprint measurement.
- Agents: what an agentic harness adds to an LLM (tools, memory, RAG, MCP), and how to develop with agents without blowing up the bill.
- Evaluation: French-language quality benchmarks, protected against contamination and statistically analysed, to pick a model on measurements rather than public leaderboards.
| Project | Status | Description |
|---|---|---|
| Agent Cost Kit | 🟢 Deployed · v0.21.1 | Claude Code plugin for developing with agents without blowing up the bill: role-routed agents, quick end-of-story check, real cost measurement from local transcripts. |
| Claude Code Mods | 🟢 Deployed · v1.0.1 | Mods for the Claude Code interface, installed with one line: diff-replie folds file diffs and shell output into one line and groups consecutive calls of the same kind, keeping the full detail one command away. |
| Generic Design System | 🟢 Deployed · v1.0.0 | Dark-first design system shipped as a Claude Code plugin: DTCG tokens, theme files for web, Streamlit, Python charts, Office and mermaid, accessibility checks (contrast, colour blindness), and a skill that audits a repository and migrates it to the system. |
| HarnessLab | 🟢 Deployed · v2.1.1 | Teaching demonstrator that shows, building block by building block, what an agentic harness adds to a bare LLM (reasoning, memory, tools, RAG, MCP, skills, hooks, sub-agents). Runs locally on CPU; French, English and German interface. |
| SLM Quality Bench | 🟢 Deployed · v0.3.5 | Hard quality benchmark for small local language models (Ollama), in French: two reasoning tracks, anti-contamination, calibrated judge, paired statistics. |
| Frugal AI Workbench | 🟠 Major rework · v1.3.0 · v2.0.0a10 pending manual testing | Local and responsible AI workbench: small language model (SLM) demo, advanced RAG and autonomous agents, with Green IT impact measurement (CodeCarbon). |
| Coding agents watch | 🔵 Built, in testing | Tooled market watch on coding agents (Claude Code first): plugins, skills and practices grouped by the problem they solve, compared with native features and checked against research papers and practitioner feedback, with a visual radar and a sourced report. |
| DevSecOps Dojo | 🔵 Built, in testing | Local teaching app to learn git and GitHub, then the CI/CD chain and DevSecOps, on real offline git with a simulated GitHub; built for vibe coders who need to understand what their agent does to their repository. |
| Gift wheel | 🔵 Built, in testing | Single-file web app (one offline HTML page under 200 KB) that spins a fair wheel to decide who gets a gift: cryptographic draw decided before the animation, accessible, works from file://. Version 1.1 (in development) adds an optional on-device chat to drive the wheel in plain French, through a rules-first cascade backed by small models running locally (decision model and SLM with Ollama); the models only call validated actions and never pick the winner. |
| Jira Copilot | 🔵 Built, in testing | Automated reporting and AI copilot for an IT migration factory driven by Jira (fictitious data). |
| LibrarianBench | 🔵 Built, in testing | Benchmark of AI document retrieval architectures (naive and hybrid RAG, GraphRAG, LLM wiki, agentic search) on quality, cost, latency, energy and compliance. |
| NN Weight Writing | 🔵 Built, in testing | Having an LLM write every weight of a small neural network directly, without training, then comparing it with hand-written rules, standard training, weight editing and activation steering, from rule-based tasks up to grafting written circuits into a small language model. |
| PentestScribe | 🔵 Built, in testing | Offline AI assistant for cybersecurity audit consultants on a 16 GB CPU-only laptop: a local small language model drafts vulnerability findings from raw pentest notes, with no data leaving the machine. |
| SLM Service Public | 🔵 Built, in testing | Offline AI assistant for French public sector staff on a 16 GB office PC: fine-tuned small model, local RAG and the official data.gouv MCP server. |
| SLM Weight Editing | 🔵 Built, in testing | Fixing small language model defects such as reasoning loops by having Claude edit their weights directly, without training, then measuring gains and regressions on SLM Quality Bench. |
| Mood Steering SLM | 🟣 Built, training or benchmarking in progress | Giving a small language model a "mood" such as joy by activation steering at inference time: dosable and reversible, with no weight change, measured against a role-play prompt on tone, accuracy, sycophancy and energy. |
| Project cockpit | 🟡 In development | Local, personal dashboard that tracks projects and parallel Claude Code sessions across several PCs: session hooks and a per-PC scan feed a local web app that shows the action to take on each project (stable, to stabilise, paused, to restart, to file) and what to resume after a shutdown. |
| EvalDojo | ⚪ Specification | Teaching app to learn how to evaluate generative AI in practice: evaluation pipelines, LLM-as-a-judge, evaluation datasets, RAG and AI search evaluation, and data and evaluation for LLM fine-tuning. |
Table regenerated weekly from the repositories themselves. Projects are documented in French. "Pending manual testing" marks a version that is published but not yet manually tested.
🇫🇷 Version française
Senior Consultant IA chez Wavestone · LinkedIn
Je construis des outils d'IA qui tournent sur des machines ordinaires, je mesure ce qu'ils valent vraiment, et je rends visible la façon dont ils fonctionnent.
- IA locale et frugale : petits modèles de langage (SLM) sur un simple PC, sans GPU ni dépendance au cloud, avec mesure de l'impact carbone.
- Agents : ce qu'un harnais agentique ajoute à un LLM (outils, mémoire, RAG, MCP), et comment développer par agents sans faire exploser la facture.
- Évaluation : bancs de qualité en français, protégés de la contamination et analysés statistiquement, pour choisir un modèle sur des mesures plutôt que sur des classements publics.
| Projet | État | Description |
|---|---|---|
| Agent Cost Kit | 🟢 Déployé · v0.21.1 | Plugin Claude Code pour développer par agents sans faire exploser la facture : agents routés par rôle, contrôle rapide, mesure du coût réel depuis les transcriptions. |
| Charte générique | 🟢 Déployé · v1.0.0 | Generic dark-first design system: DTCG tokens, theme files (CSS, Python, Streamlit, Office, mermaid) and accessibility checks, as a Claude Code plugin. |
| Claude Code Mods | 🟢 Déployé · v1.0.1 | Mods pour l'interface de Claude Code, installables en une ligne |
| HarnessLab | 🟢 Déployé · v2.1.1 | Démonstrateur pédagogique qui montre, brique par brique, ce qu'un harnais agentique ajoute à un LLM nu (raisonnement, mémoire, outils, RAG, MCP, skills, hooks, sous-agents). Local, sur CPU ; interface en français, anglais et allemand. |
| SLM Quality Bench | 🟢 Déployé · v0.3.5 | Banc de qualité difficile pour petits modèles de langage locaux (Ollama), en français : deux pistes de réflexion, anti-contamination, juge calibré, statistiques appariées. |
| Frugal AI Workbench | 🟠 En refonte majeure · v1.3.0 · v2.0.0a10 à tester | Workbench d'IA locale & responsable : démo de Small Language Models (SLM), RAG avancé et agents autonomes avec mesure d'impact Green IT (CodeCarbon). |
| Copilote Jira | 🔵 Développé, en test | Reporting automatisé et copilote IA d'une usine de migration de SI depuis Jira (données fictives) |
| DevSecOps Dojo | 🔵 Développé, en test | Application pédagogique locale pour apprendre git et GitHub, puis la chaîne CI/CD et DevSecOps, sur du vrai git hors ligne avec un GitHub simulé ; pensée pour les vibe coders. v0.1.0 en test. |
| LibrarianBench | 🔵 Développé, en test | Banc de comparaison des architectures de recherche documentaire IA (RAG naïf, hybride, GraphRAG, LLM Wiki, recherche agentique…) sur la qualité, le coût, la latence, l'énergie et la conformité. |
| NN Weight Writing | 🔵 Développé, en test | Having an LLM write every weight of a small neural network directly, without training, compared with hand-written rules, standard training, weight editing and activation steering |
| PentestScribe | 🔵 Développé, en test | Assistant IA hors ligne pour les consultants en audit de cybersécurité, sur un PC de 16 Go sans GPU : rédaction de fiches de vulnérabilité à partir des notes de pentest par un petit modèle local, sans qu'aucune donnée ne quitte le poste. |
| Roue du hasard | 🔵 Développé, en test | Roue du hasard : une page HTML autonome pour désigner qui reçoit un cadeau |
| SLM Service Public | 🔵 Développé, en test | Assistant IA hors ligne pour les agents du secteur public, sur un PC bureautique de 16 Go : petit modèle fine-tuné, RAG local et serveur MCP officiel de data.gouv. |
| SLM Weight Editing | 🔵 Développé, en test | Fixing small language model defects (reasoning loops…) by having an LLM (Claude Opus in Claude Code) edit their weights directly, without training — measured against the original by slm-quality-bench |
| Veille agents de code | 🔵 Développé, en test | Veille outillée sur les agents de code (Claude Code) : plugins, skills et pratiques classés par problème, vérifiés contre la recherche et les praticiens |
| Mood Steering SLM | 🟣 Développé, entraînement ou bancs d'essai en cours | Giving a small language model a "mood" (joy, calm…) by activation steering at inference time — dosable, reversible, no weight change — compared with a role-play prompt on tone, accuracy, sycophancy and energy |
| Pilotage projets | 🟡 En développement | Tableau de bord local : suit les projets et les sessions Claude Code menées en parallèle sur plusieurs PC, et indique quoi relancer à la reprise. |
| EvalDojo | ⚪ En spécification | Application pédagogique pour apprendre à évaluer l'IA générative en pratique : pipelines d'évaluation, LLM juges, jeux de données, RAG et fine-tuning |
Tableau régénéré chaque semaine à partir des dépôts eux-mêmes. « À tester » désigne une version publiée mais pas encore recettée manuellement.