TAMSI

2026 · MCP & AI devtools

AI Engineer — MCP agents & AI workflows

Based in Paris — I build MCP servers, LLM workflows, and Next.js apps for product teams and enterprise clients.

About

"“Controlling complexity is the essence of computer programming.”" — Brian Kernighan

AI Engineer in Paris with ~8 years of experience at Livingcolor, contracting for TotalEnergies, France Télévisions, AFP, TV5Monde, Harmonie Mutuelle, and LVMH houses (Krug, Moët, Hennessy).

On the side I ship open source tools on GitHub — HuggiMon, livingcolor-skills, redbee-mcp — and contribute to products like VisualQ (visual regression testing). Trained at 42, MCP & AI Agents certified (Hugging Face) and ML specialization (Coursera).

location: Paris, France

What I focus on

Aligned with my GitHub profile and recent open source work.

MCP

Model Context Protocol servers for Cursor, Claude Desktop, and IDEs — bridges to business APIs and code review tools.

AI & code quality

Automated review, tech debt, security, and test gaps — LLM workflows (OpenAI-compatible, Ollama, Hugging Face).

Web product

Next.js, TypeScript, Firebase, Playwright — from enterprise CMS (Drupal) to React Native apps and SaaS platforms.

Languages

French: Native

English: Professional

Spanish: Elementary

Experience & Education

AI & Web Engineer

org: Livingcolor

period: Jan 2018 — Present

location: Lieusaint, Île-de-France

Drupal & Symfony sites, React / Next.js apps (France TV, Harmonie Mutuelle, TV5Monde), React Native (Asmodee, Crédit Agricole), Shopify, and AI integrations (MCP, Playwright, Firebase) for team workflows.

AI Agents Fundamentals

org: Hugging Face

period: Jun 2026

location: Online

Certification on AI agent fundamentals — tool calling, memory, orchestration, and deploying autonomous agents.

Fundamentals of MCP

org: Hugging Face

period: Nov 2025

location: Online

Certification on the Model Context Protocol — wiring AI tools into IDEs and agents.

Machine Learning Specialization

org: Stanford Online · Coursera

period: May 2024

location: Online

Supervised & unsupervised learning, neural networks, recommender systems, reinforcement learning.

Level 11 Certificate — Computer Programming

org: 42

period: 2021

location: Paris

Intensive C programming, algorithms, and collaborative projects.

Interests & tooling

What shapes my day-to-day — agentic IDEs, local LLMs, MCP, plugins, and automation.

Cursor & agents

Primary IDE — rules, plugins, skills, and agent sessions to code, review, and iterate faster.

• Cursor
• Rules
• Plugins
• Agent skills
• Composer

demo_video: https://tamsi.dev/interests/videos/cursor.mp4

LLM & inference

Local and remote models — offline prototyping, Hugging Face demos, OpenAI-compatible endpoints.

• Ollama
• Hugging Face
• OpenAI-compatible
• Qwen
• Gradio

demo_video: https://tamsi.dev/interests/videos/ollama.mp4

MCP & protocols

Model Context Protocol — custom servers exposing GitHub, business APIs, and code review to assistants.

• MCP
• MCP servers
• Tool calling
• Claude Desktop
• Cursor MCP

demo_video: https://tamsi.dev/interests/videos/mcp.mp4

Automation

Cutting repetitive work — CI, visual testing, scripts, and pipelines around repos.

• Playwright
• GitHub Actions
• GitLab CI
• Scripts & hooks
• Vercel

demo_video: https://tamsi.dev/interests/videos/playwright.mp4

Hermes Agent

Open-source agent (Nous Research) with a learning loop — persistent skills, memory, and cross-platform automation.

• Hermes Agent
• Skills
• Persistent memory
• Webhooks
• Firebase

demo_video: https://tamsi.dev/interests/videos/hermes.mp4

Projects

Curated from my GitHub — live demos when available.

[Featured] huggimon

stack: TypeScript

Pokémon TCG-style AI trainer card from your Hugging Face profile — 3D tilt, holo shaders, a dedicated page per username.

[Featured] livingcolor-plugin

stack: Python · Hermes

Hermes Agent plugin for autonomous delivery — work orders, Jira gates, and human-approved deploys.

[Featured] livingcolor-skills

stack: TypeScript

Portable, versioned Hermes skills — reusable expert behaviors for AI agents.

[Featured] redbee-mcp

stack: Python · MCP

MCP bridge to Red Bee Media OTT APIs — for Cursor, Claude, and MCP clients automating the streaming platform.

[Featured] visualq-mcp

stack: TypeScript · MCP

VisualQ MCP bridge — run VRT, poll failures, and approve baselines from Cursor or Claude.

ai-code-reviewer-mcp

stack: TypeScript · MCP

MCP server for GitHub code review — bugs, security, tech debt, missing tests via an OpenAI-compatible LLM. Gradio demo on Hugging Face.

git-mentor

stack: TypeScript · CLI

Local-first GitHub career coach — Ink TUI, Ollama, evidence-backed recommendations from your repo history.

handship

stack: Python · Gradio

Hand-controlled spaceship game — webcam and pose detection, playable directly on Hugging Face.

livingcolor-evolution

stack: TypeScript · Hermes

Auto-evolution of Hermes skills from the web — continuous audit and refresh of agent behaviors.

Blog

Why first — then the essentials of how (stack, quantization, MCP).

Jev — classify without paying an LLM every turn

date: 2026-09-18

TypeSafe ships Jev, a System One model: typed decisions, no text. LangChain turns it into a harness classifier — routing and auto-mode — at a fraction of chat cost.

Ternary Bonsai 2 27B — Qwen3.8 in 6 GB, 98% of FP16

date: 2026-09-18

PrismML recompresses Qwen3.8-27B into true ternary weights (1.72 bits). 5.95 GB, 98.2% of FP16, ~47 tok/s on an M5 Max. The catch: it will not run in stock llama.cpp.

Hermes HUD — play WoW and prompt at the same time

date: 2026-08-25

Brooklyn (Nous Research) drops the agent as an overlay on a WoW session. The HUD is no longer a window you alt-tab to: it’s a layer, and its position is the context.

DFlash 2 — 3× more tokens on Qwen3.8-27B, same output

date: 2026-08-21

Inco AI’s parallel drafter for Qwen3.8-27B: +20% acceptance vs DFlash, 2.7–3.4× autoregressive throughput, identical output. Why it matters as soon as an agent loops.

Qwen 3.8 27B — the open 27B that pulls the frontier onto local hardware

date: 2026-08-18

Qwen3.8-27B specs, official scores versus 3.6 and Opus 4.6 Max, and what that changes for local AI.

Unsloth Studio — run and train LLMs locally

date: 2026-08-12

Studio, Desktop, and Unsloth kernels: GGUF / MLX, low-VRAM fine-tunes, Data Recipes — and why the Qwen3.8-27B GGUF fits in 17 GB.

HuggiMon — your AI trainer card from your Hugging Face profile

date: 2026-07-09

A Next.js site that turns public Hub activity into an interactive Pokémon TCG card — holo shaders, follower binder, PNG export, and a shareable profile page.

Unsloth Studio HF live — train and run LLMs locally, no cloud

date: 2026-06-27

Notes from the Hugging Face live with Daniel Hanchen (UnslothAI): Studio, dynamic quants, benchmarks, and low-VRAM fine-tuning on Mac/Windows/Linux.

Why I’m moving to Hermes — automate with cheaper models

date: 2026-06-09

Fable 5 just landed and tops the Cursor benchmark at a steep price: I’m shifting agentic work to Hermes and affordable models.

Why I run a remote Qwen 3.6 27B server instead of paying Cursor

date: 2026-06-01

Free tokens instead of Cursor models — Qwen 3.6 27B on AWS via vLLM (4-bit), Ollama locally for git-mentor.

Why an MCP code reviewer — not just the Cursor chat

date: 2026-05-15

Give the agent concrete GitHub actions instead of one big “look at my repo” prompt.

redbee-mcp: testing OTT APIs day-to-day with my AI

date: 2026-05-01

Built in a hurry when MCP landed — so Cursor could call the real Red Bee Media APIs I need at work.

Contact

Interested in MCP, AI tooling, or a web product? Drop me a line.

© 2026 Tamsi Besson · tamsi.dev