Mickael Gicquaire

AI Solutions Developer

  • Haute-Saône (France) — moving closer to the Swiss border once hired
  • French national (EU) — eligible for a Swiss G permit
  • Available immediately — 100% workload
Phone
+33 6 40 51 77 44
LinkedIn
linkedin.com/in/mickael-gicquaire
Web CV
cv.mickael-gicquaire.com
Photo of Mickael Gicquaire

Profile

In personal projects, I design and evaluate hybrid RAG systems and MCP servers (golden sets, automated tests), using agent-driven development. Previously Technical & Marketing Manager at a publishing house in Lausanne: websites, sales funnels, Salesforce Marketing Cloud campaigns, AI tools for automation and analytics. 6+ years as a digital entrepreneur. Art-school background: storytelling and AI visuals.

Working method: need, design, AI build, measure, iterate; loop from measurement back to design.

Key figures

  • 800,000+passages indexed locally (2 hybrid RAG systems)734,456 + 69,034 passages
  • 4 languages · 26 corporalocal multilingual search engine (734,456 passages)

Professional experience

Applied AI R&D — self-directed projects

09/2025 – present
  • Designed, specified, evaluated and optimised each system; agent-driven development (Claude Code, adversarial review by Codex), run locally.
  • Indexing and search run locally on a 2014 laptop (Intel i7-4860HQ, GTX 980M 4 GB GPU) under Windows + WSL2.
  • Independent activity (sole proprietorship): digital educational resources; web and AI development.

Projects detailed on page 2.

Technical & Marketing Manager

09/2024 – 09/2025

Les éditions Nouvelle Page (Vivara group) — permanent contract, Lausanne

Swiss publishing house with multi-million revenue; reported directly to the Publisher (who runs the imprint).

  • Designed and built websites and sales funnels (WordPress, PHP) and in-house funnel-building tools (JavaScript, HTML, CSS).
  • Email campaigns and sequences in Salesforce Marketing Cloud; order bumps and conversion optimisation.
  • Introduced AI tools for automation and analysis.

Founder — online training & digital marketing business

2017 – 04/2024

Health and wellness sector.

  • Designed and built 50+ landing pages and websites; end-to-end sales funnels (opt-in pages, email sequences).
  • Owned offer, positioning, acquisition and customer relationship end to end.
Mickael Gicquaire — page 1/2

AI projects

Hybrid RAG over technical documentation — “atelier-code” MCP server

STACKPython, FastMCP, ChromaDB, sentence-transformers, SQLite FTS5 (BM25), RRF

  • MCP server (FastMCP, 8 tools) for hybrid BM25 + vector search fused with RRF, over ~69,000 passages from 31 technical books, integrated with Claude Code.
Protocol
A/B test of the answer prompt on a 61-question golden set, LLM judge
Metric
internal score “answer judged correct + citation present”, over the 56 questions covered by the corpus
Result
Minimal instruction (baseline)21/56 (38%)Answer protocol51/56 (91%)

5/5 out-of-corpus abstentions in both modes

Limitation
the gain comes from citations, not accuracy; uncalibrated LLM judge

Multilingual hybrid RAG engine over a heritage corpus (ancient texts) — MCP server

STACKPython, FastMCP, multilingual sentence-transformers, PyTorch CUDA, NumPy float16 (memory-mapped), SQLite FTS5, RRF

  • Search engine exposed as an MCP server (6 tools): 734,456 passages, 26 corpora, 4 languages, 2 of them in non-Latin scripts; search runs fully locally, answers generated by Claude.
  • Migrated the ChromaDB file (float32 vectors, passage text and internal index: 18.8 GB) to memory-mapped float16 NumPy vectors only (1.1 GB), with text now served from the existing full-text index.

Local voice assistant for a coding agent

STACKPython, FastAPI, WebSocket, MCP, Claude Code hooks, faster-whisper, Piper TTS, pytest

  • Local voice layer for a coding agent: Claude Code replies spoken through Piper synthesis (MCP server and hooks); speech → faster-whisper transcription → voice loop validated as a prototype.
  • Speech synthesis: ~263 ms median from text received to first audio (n=5, excluding WSL↔Windows transit); engines chosen by benchmark with thresholds fixed beforehand; 74 pytest tests (phase 1).

Traceable, source-backed historical research platform

  • Every published fact links to a captured source sealed with a SHA-256 hash: 1,849 citations, each checked by a verification agent.
  • Specialised agents (research, source checking, fidelity control) under automated validation: 16 JSON schemas, 113 test cases, pre-commit hook.

Multi-agent long-form writing assistant

  • Multi-agent workflow (Claude drafting, Codex as read-only adversarial reviewer): roles, write permissions and human sign-off set by contract, over 150,000 words.

Skills

Applied AI
Hybrid RAG: ChromaDB, SQLite FTS5 / BM25, RRF fusion, sentence-transformers, PyTorch CUDA · MCP: FastMCP, servers and tools, Claude Code integration · Coding agents: Claude Code, Codex; evaluation with golden sets and LLM judges
Development
Python (through agent-driven development and code reading) · C (solid) · JavaScript, HTML, CSS, PHP, WordPress · React, Node.js (basics) · Three.js, GSAP, Lenis
Digital marketing
Salesforce Marketing Cloud, SEO/GEO, Meta Ads, copywriting, sales funnels
Creative AI
Midjourney, Higgsfield (proficient) · ComfyUI, LoRA training (currently learning) · ElevenLabs (basics)
Systems & tools
Linux, WSL2, Git · n8n (basics)

Education

  • Self-study: App Academy Open, The Odin Project, self-taught 42 “piscine” curriculum2024
  • Isart Digital — Sandbox 2014: one-month intensive selection (bootcamp “piscine” format), ranked 1st08/2014 – 09/2014
  • École Émile Cohl — coursework in graphic design (partial programme)09/2009 – 06/2010
  • ArtFX — coursework in 3D modelling, animation and visual effects (partial programme)09/2007 – 06/2009

Languages

  • French — native ·
  • English — B2 (self-assessed)

References

References available on request