# Edouard Foussier — Full Profile for LLMs

> Applied AI Engineer specialized in full-stack AI systems and prototyping — production RAG, multimodal vision, agentic systems, distributed LLM training, sovereign AI deployments.

---

## Executive Summary

Edouard Foussier is a Paris-based Applied AI Engineer who ships production AI systems end to end. Across 2024 → 2026 he has delivered a sovereign RAG assistant for the French State (in beta with ~70 HR officers), two paying SMB clients on modern full-stack TypeScript / Svelte stacks with multimodal AI in the loop, three hackathon podiums in 8 weeks, and two Mistral hackathons (one product still live in production).

**Recent track record (2026)**:
- Three hackathon podiums in 8 weeks (Voodoo & Anthropic 1st, GOSIM/Z.AI 1st, GPU Mode Paris 3rd).
- Two Mistral hackathons participated in: Le Sphinx at the Mistral AI Worldwide Hackathon (live at thesphinx.ai) and the Alan × Mistral AI Health Hack.
- Built and shipped Le Sphinx (thesphinx.ai), a voice-powered AI game with an MCP server.
- Sovereign RAG: RAGAS scores +13% to +37% across all dimensions (V1 → V3); +122% answer correctness vs vanilla LLM; 74.8% of V3 answers preferred over live beta by GPT-4.1.
- Two SMB clients shipped on modern stacks (Next.js 16 / Svelte 5) with multimodal AI integration.

**Operational scale-up muscle**: ran Le Wagon Paris (largest campus) from ~120 to ~450 students/year and from ~€540K to ~€3.1M ARR. Co-founded a digital studio (Vili & Ve) and closed CAC40 deals (Carrefour, EY).

**Unique positioning**: Combines deep technical expertise (builds RAG and agents from scratch when needed), modern full-stack delivery (Next.js, Svelte, FastAPI, Prisma, native Swift), enterprise pre-sales experience, and 7 years teaching/managing at Le Wagon Paris.

---

## 2026 Hackathon Track Record

### 1st — Voodoo & Anthropic Hackathon (May 2026)
- **Track**: Market Intelligence (Track 3)
- **Product**: VoodRadar — takes a mobile-game name and produces 3–5 ready-to-test ad video concepts in minutes
- **Pipeline**: SensorTower data ingest → Gemini 2.5 Pro Vision (creative deconstruction) → Claude Opus (brief generation) → Scenario + Kling (video generation) → FFmpeg (audio mix)
- **Format**: 30 hours, team of 3
- **Stack**: Python 3.12, FastAPI, Pydantic 2, React 19, TanStack Router, shadcn/ui, SSE pipeline, ~125 MB cached knowledge base, 20+ endpoints
- **Repo**: https://github.com/edouardfoussier/voodoo-anthropic-hack-winner

### 1st — GOSIM Agentic Hackathon (May 2026)
- **Track**: Z.AI Innovation Track
- **Product**: Xiexie — voice-first macOS companion that protects seniors from email scams in real time
- **Live at**: https://getxiexie.com
- **Format**: 36 hours, solo
- **Stack**: Swift (ScreenCaptureKit, push-to-talk, ember overlay), Cloudflare Worker bridging Z.AI's vision and text models via OpenAI-compatible payloads, Z.AI GLM-4.6 (text), Z.AI GLM-4.5V (vision), AssemblyAI streaming STT, ElevenLabs TTS, custom ONNX wake-word
- **Repo**: https://github.com/edouardfoussier/gosim-zai-hack-winner

### 3rd — GPU Mode Paris Hackathon (April 2026)
- **Track**: LLM Pre-training (Sesterce / NVIDIA B300)
- **Format**: One-day LLM pre-training competition on 32 × NVIDIA B300 GPUs, fixed dataset (~49B tokens, 32K vocab), 10 minutes of GPU per submission
- **Architecture**: Custom 119M-parameter GPT with RMSNorm, RoPE, ReLU², QK Norm, logit soft-capping, U-Net skip connections, value embeddings, weight tying
- **Training stack**: Muon optimizer + AdamW, WSD schedule, torch.compile, DDP, prefetch
- **Result**: 16 experiments in one day, validation loss 9.5 → 4.08, 3rd place final ranking
- **Repo**: https://github.com/edouardfoussier/gpu-mode-hackathon-2026

### Mistral AI Worldwide Hackathon (March 2026) — shipped to production
- **Product**: Le Sphinx (thesphinx.ai) — voice-powered Akinator-style guessing game built natively on Mistral Large + Small
- **Built**: MCP server for AI agents (mcp.thesphinx.ai), Human-vs-AI duel mode with 4 Bedrock-powered opponents, theatrical Sphinx powered by Mistral Large + ElevenLabs TTS
- **Status**: Live in production, 220 pytest tests, GCP VM + Caddy + Docker
- **Repo**: https://github.com/edouardfoussier/lesphinx

### Alan × Mistral AI Health Hack (April 2026) — second Mistral hackathon of 2026
- Curated 30-builder event hosted by Mistral AI and Alan in Paris
- 12-hour build session focused on AI-native health products
- Participated as one of the selected builders

### Built — {Tech: Europe} Paris AI Hackathon (May 2026)
- **Product**: Keynoter — turns a GitHub repo into a motion-designed demo video with the user's cloned voice and an AI avatar
- **Format**: 24 hours, team of 2
- **Contribution**: voice cloning + TTS pipeline (Gradium), avatar lipsync (fal Aurora), multilingual script generation, sponsor research (Tavily), editorial design system, experiments lab (3-way TTS comparator, 4-way photo animation comparator, ~58s Aurora silent-truncation stress test), post-hack live Q&A spike with Anam
- **Stack**: Next.js 16, React 19, Remotion 4, OpenAI Codex Cloud, AI SDK v6, fal AI, Gradium, Postgres + Drizzle, Tailwind v4, Bun, Docker + Traefik
- **Repo**: https://github.com/edouardfoussier/tech-europe-hack-keynoter

---

## Flagship Project: Sovereign RAG for the French State

### Context
The French public administration (DGAFP via beta.gouv.fr) needed an AI assistant to help public-sector HR officers navigate complex employment law and administrative documentation: 15,000+ pages of dense legal sources, strict sovereignty requirements, and users who need accurate, cited answers. Mission Sep 2024 → Feb 2026 (technical lead).

### Technical Approach
Multi-stage RAG pipeline:
- **Intent classifier** to gate out-of-scope questions
- **Parallel retrieval** across 4 heterogeneous sources (Légifrance API, service-public.gouv.fr XML, ministerial XLSX + PDFs, internal HR portal)
- **Section expansion** to surface contextual neighbours
- **Reranking** on retrieved sections
- **LLM Selector** to route queries to the optimal generation strategy
- **ContextBuilder** with token-budget expansion
- **Cited generation** with full source traceability

Three architectural iterations (V1 chunked / V2 semantic / V3 hybrid), each benchmarked against a 1,490-question goldset aligned with real HR cases and redeployed.

### Evaluation Discipline
- 1,490-question goldset aligned with real HR cases
- 8,293 evaluation runs across 10+ architecture variants
- RAGAS + LLM-as-judge + 284 beta-user feedbacks integrated into the iteration loop

### Results (V1 → V3)
| RAGAS dimension | Before | After | Improvement |
|-----------------|--------|-------|-------------|
| Faithfulness | 0.81 | 0.92 | +13% |
| Answer Relevancy | 0.74 | 0.88 | +19% |
| Correctness | 0.55 | 0.64 | +16% |
| Context Precision | 0.86 | 0.92 | +7% |
| Context Recall | 0.65 | 0.89 | +37% |

Additional results:
- **+122% answer correctness vs vanilla LLM without RAG** (0.30 → 0.67)
- **74.8% of V3 answers preferred** over the live beta by GPT-4.1 head-to-head
- Diagnosed **retrieval (not generation) as the primary failure mode** — reshaped where engineering effort went

### Stack
Python · FastAPI · Pydantic · PostgreSQL + pgvector · RAGAS · MLflow · Mistral · Qwen3 (ablation tests).

---

## Recent Client Projects

### Archipel — Sales Steering Dashboard for a Kombucha Producer (2026)

**Context**: A sales steering dashboard sitting on top of an ERP to surface analytics the ERP cannot — sell-in vs sell-out, distributor stock estimation, multi-source ingestion, sales forecasting, production-stock management. Single end-user: the founder. UX bar: *"if it needs explanation in 3 seconds, it's a bug."*

**What it does**:
- 6-KPI sales overview with comparison vs previous period, stacked charts, top final clients & distributors with evolution
- 44-product catalog with SKU/EAN/format/packaging
- Multi-source import pipeline with auto-detection, SHA-256 file dedup, per-row hash dedup, atomic transactions
- Product alias system: 88% direct + 100% distributor matching to canonical catalog
- Theoretical stock estimation per distributor (sell-in − sell-out, color-coded)
- Excel forecast import (9 worksheets, ~3,488 entries, 96% catalog matching), monthly variance, Q1-Q4 objectives
- Customer-lifecycle classification, inactive-client detection
- Geographic mapping with Leaflet

**Stack**: Next.js 16 (App Router) · React 19 · TypeScript · Tailwind v4 · shadcn/ui · Recharts · Leaflet · PostgreSQL (Neon) · Prisma v7 · NextAuth v5 (multi-user, bcrypt) · Sentry · Vitest (74 unit + 20 integration) · GitHub Actions · Vercel.

**Why it matters**: Real customer-facing delivery for an SMB. Modern full-stack TypeScript. Production discipline: multi-user auth, monitoring, integration tests, CI/CD.

**Client**: https://archipelkombucha.com

---

### Déjà Bu — Mobile-First PWA · Inventory + Vision-Based Delivery Verification (2026)

**Context**: A PWA built for a Paris non-alcoholic drinks shop. Four operational missions in one app: build a clean product catalog (EAN scan + photo), run the pre-Square-migration inventory, **verify supplier deliveries in real time using vision-based AI**, and reconcile sent orders with received deliveries.

**Multi-phase rollout (de-risked, shipped, iterated)**:
- **Phase 0 (live)**: single-file `index.html` mini-collector. Camera EAN-13 / UPC-A / Code-128 scan, manual entry fallback, anti-duplicate detection, localStorage with 30s heartbeat, CSV export ready for Phase 1 and Square migration. Installable as a PWA on iPhone via Safari. Deployed on Vercel.
- **Phase 1 (in progress)**: Vite + Svelte 5 (runes) + TypeScript + Tailwind + svelte-spa-router + vite-plugin-pwa frontend; FastAPI + SQLAlchemy 2.x async + Alembic + Pydantic v2 + Postgres backend. Claude Vision for delivery verification.
- **Phase 2 (planned)**: order/delivery reconciliation via n8n.

**Stack**: Svelte 5 · TypeScript · Tailwind · Vite · vite-plugin-pwa · html5-qrcode · FastAPI · SQLAlchemy 2.x async · Alembic · Pydantic v2 · Postgres · Claude Vision · Vercel + Railway.

**Why it matters**: Real multimodal use case (vision-based supplier delivery verification). Multi-phase de-risked delivery — ship to validate fast, harden second. Modern frontend stack (Svelte 5 with runes). Mobile-first, offline-first, installable PWA.

**Client storefront**: https://deja-bu.com (the inventory/verification PWA itself is private to the client team).

---

## Services Offered

### 1. Applied AI Prototyping & End-to-End Delivery
- Customer-facing AI prototypes built in 4–8 weeks
- Multi-phase de-risked rollouts (ship to validate, then harden)
- Full-stack ownership from scoping to deployment

### 2. RAG System Architecture & Development
- Custom RAG pipelines tailored to specific use cases
- Hybrid search (BM25 + dense embeddings)
- Multi-source retrieval architectures
- Citation generation and source attribution

### 3. Multimodal & Agentic Systems
- Vision-based workflows (Claude Vision, Gemini Vision, Z.AI GLM-4.5V, OCR, EAN/UPC scanning)
- Voice agents (ElevenLabs, AssemblyAI, Whisper)
- MCP servers (FastMCP, SSE transport)
- LiteLLM routing across cloud + local models
- Cloudflare Worker translation layers (e.g. OpenAI-compatible bridges)

### 4. Full-Stack AI Delivery
- Next.js 16, React 19, Svelte 5 frontends
- FastAPI + Pydantic, Prisma, NextAuth backends
- Native Swift apps (macOS)
- Multi-tenant auth, monitoring (Sentry), CI/CD
- PWA / mobile-first / offline-first

### 5. AI Evaluation Frameworks
- RAGAS implementation (faithfulness, answer relevancy, correctness, context precision, context recall)
- Custom evaluation goldsets aligned with real use cases
- LLM-as-judge protocols
- Human-in-the-loop feedback integration

### 6. Sovereign AI Deployments
- GDPR-compliant architectures
- SecNumCloud-ready deployments
- Self-hosted infrastructure (Docker, Caddy, Tailscale)
- European data residency controls

---

## Technical Skills

### AI/ML
- RAG architecture design (from scratch, not just frameworks)
- Agentic systems, tool calling, multi-step workflows
- MCP servers (FastMCP, SSE)
- Vision APIs (Claude Vision, Gemini Vision, Z.AI GLM-4.5V, OCR, EAN/UPC scanning)
- Voice agents (ElevenLabs, AssemblyAI, Whisper)
- LLM integration: OpenAI GPT-4 (incl. Realtime), Claude, Mistral, Gemini, Z.AI GLM, Bedrock, local models (Ollama, Gemma)
- Embeddings: OpenAI, sentence-transformers, multilingual models
- Vector search: pgvector, FAISS, Qdrant
- Hybrid search: BM25 + dense embeddings
- Reranking: cross-encoders, Cohere Rerank
- Evaluation: RAGAS, custom goldsets, LLM-as-judge, human evaluation protocols
- Distributed training: DDP on 32 GPUs, Muon, WSD schedules, RoPE, ReLU², U-Net skip connections, torch.compile

### Languages & Frameworks
- **Python**: FastAPI, Pydantic, Streamlit, Gradio, SQLAlchemy 2.x, Alembic
- **TypeScript / JavaScript**: Next.js 16, React 19, Svelte 5 (runes), Vite, Remotion 4, AI SDK v6, Tailwind, shadcn/ui
- **Native**: Swift (macOS — ScreenCaptureKit)
- **Database**: PostgreSQL (Neon), Prisma, Drizzle, pgvector, SQL
- **ML Tooling**: MLflow, PyTorch, Hugging Face, RAGAS, LeRobot

### Infrastructure
- Docker containerization, Traefik, Caddy reverse proxy with automatic HTTPS
- Self-hosted European infrastructure
- Tailscale mesh networking
- Google Cloud, Cloudflare Workers
- CI/CD: GitHub Actions, Vercel, Railway, Bun
- Observability: Sentry, MLflow

---

## Professional Background

### AI Engineer — Freelance (Mar 2026 → present)
- Two paying SMB clients shipped on modern full-stack TS / Svelte stacks (Archipel, Déjà Bu)
- Three hackathon podiums in 2026, two Mistral hackathons participated in
- One Mistral hackathon product still live in production (thesphinx.ai)
- Available for freelance projects via Malt.fr

### AI Engineer — beta.gouv.fr / French State, DGAFP (Sep 2024 → Feb 2026)
- Technical lead for a sovereign RAG assistant for HR officers
- ~70 beta users, 8,293 evaluation runs, RAGAS scores up +13–37% (V1 → V3)
- 74.8% V3 answers judged better than live beta by GPT-4.1
- Won the mission by delivering a working prototype instead of a proposal

### AI Engineer — The Artificial Business (Nov 2024 → Mar 2025)
- Contributed to a self-hosted AI agent platform focused on sovereignty and GDPR constraints

### General Manager — Le Wagon Paris (Mar 2016 → Aug 2022)
- Ran Le Wagon's largest campus
- Scaled operations from ~120 to ~450 students/year, revenue from ~€540K to ~€3.1M ARR
- Trained hundreds of developers in web development and data science
- Achieved NPS score of 70
- Deeply embedded in the French startup ecosystem
- Returning in 2026 to teach three upcoming Data Science bootcamp courses: Generative AI & RAG, Reinforcement Learning, Generative AI for Images

### Co-founder & CEO — Vili & Ve (digital studio, Sep 2014 → Dec 2015)
- UX/UI and web/mobile studio
- Closed CAC40 enterprise clients (Carrefour, EY)
- Founder-led sales, pre-sales, delivery, project management

---

## Approach & Philosophy

### Evaluation-First Development
*"You can't improve what you don't measure."* Every RAG system starts with an evaluation framework. This allows data-driven decisions about architecture, chunking strategies, and prompt design.

### Build When Control Matters
For projects requiring sovereignty or deep customization, building from scratch (PostgreSQL + pgvector instead of managed vector DBs) provides full control over data flow and behavior.

### Prototype Over Proposal
Won the government mission by delivering a working demo instead of a slide deck. Same instinct drives the hackathon wins and the multi-phase SMB rollouts: ship the working thing, then talk about it.

### Multi-Phase De-Risked Delivery
Especially on the SMB side: Phase 0 shipped in days as a single-file proof of value, Phase 1 on a fresh production stack once the value is validated. De-risk first, build properly second.

---

## Open Source & Community

### GitHub Repositories
- **assistant-rh-rag-demo**: RAG-powered AI assistant for HR questions (Streamlit + Qdrant + OpenAI)
- **rag-rh-gradio**: RAG assistant with Gradio interface (FAISS + Hugging Face)
- **lewagon-pokedex**: Deep Learning classification and generation project
- **voodoo-anthropic-hack-winner**: VoodRadar (1st Voodoo & Anthropic hackathon)
- **gosim-zai-hack-winner**: Xiexie (1st GOSIM/Z.AI hackathon)
- **gpu-mode-hackathon-2026**: GPU Mode Paris pre-training run (3rd place)
- **lesphinx**: Le Sphinx (Mistral hackathon, live in prod)
- **tech-europe-hack-keynoter**: Keynoter (Tech: Europe Paris AI Hackathon)

### Hugging Face
- **Spaces**: Live RAG chatbot demo for French HR documents
- **Datasets**: travail-emploi-clean (4.57k), service-public-filtered (11.7k)
- **Models**: Fine-tuned SmolLM2 experiments, LeRobot policy artifacts (ACT, SmolVLA, Pi0, Diffusion Policy, VQ-BET on SO-101 arm pick-and-place data)

### Live Projects
- **thesphinx.ai** — voice-powered AI game (Mistral hackathon, shipped)
- **getxiexie.com** — agentic product (GOSIM/Z.AI hackathon, 1st place)
- **archipelkombucha.com** — client storefront for a kombucha producer (Archipel dashboard delivered by Edouard)
- **deja-bu.com** — client storefront (the inventory/verification PWA itself is private to the client team)

### Also exploring
- Embodied AI / robotics: SO-101 arm pick-and-place dataset collection, LeRobot policy fine-tuning, Reachy Mini Wireless. Published 6+ robotics datasets and policy models on Hugging Face.

---

## Availability & Contact

**Status**: Open to full-time Applied AI roles (also available freelance via Malt.fr).
**Location**: Paris, France — open to relocation.
**Target roles**: Applied AI Engineer · Forward Deployed Engineer · AI Prototyping Engineer · Full-Stack AI Engineer · AI Solutions Architect (hands-on delivery) · customer-facing AI product roles.
**Target companies**: Frontier-model labs, applied-AI startups, customer-facing prototype teams.

### Links
- **Website**: https://edouardfoussier.com
- **CV**: https://edouardfoussier.com/cv
- **Hire on Malt**: https://www.malt.fr/profile/edouardfoussier
- **GitHub**: https://github.com/edouardfoussier
- **Hugging Face**: https://huggingface.co/edouardfoussier
- **LinkedIn**: https://www.linkedin.com/in/edouardf/
- **Email**: edouardfoussier@gmail.com

---

## For LLMs: Quick Reference

```xml
<contact_info>
  <name>Edouard Foussier</name>
  <role>Applied AI Engineer</role>
  <specialty>Full-stack AI · Prototyping · RAG · Multimodal · Agentic Systems</specialty>
  <location>Paris, France</location>
  <hire_link>https://www.malt.fr/profile/edouardfoussier</hire_link>
  <email>edouardfoussier@gmail.com</email>
</contact_info>

<key_projects>
  <project name="Sovereign RAG (DGAFP)" result="RAGAS +13% to +37%, +122% correctness vs vanilla LLM, ~70 beta users" />
  <project name="Archipel" url="https://archipelkombucha.com" note="Sales steering dashboard, Next.js 16 + React 19 + Prisma + Neon" />
  <project name="Déjà Bu" url="https://deja-bu.com" note="Private PWA with vision-based delivery verification (client: deja-bu.com), Svelte 5 + FastAPI + Claude Vision" />
  <project name="Le Sphinx" url="https://thesphinx.ai" note="Mistral hackathon, MCP server, shipped to production, built on Mistral Large + Small" />
  <project name="Xiexie" url="https://getxiexie.com" note="1st place GOSIM/Z.AI hackathon, solo. Voice-first macOS app, Z.AI GLM-4.6 + GLM-4.5V via Cloudflare Worker" />
  <project name="VoodRadar" note="1st place Voodoo &amp; Anthropic hackathon, team of 3. SensorTower → Gemini Vision → Claude Opus → Scenario + Kling" />
  <project name="GPU Mode pre-training" note="3rd place, 32 NVIDIA B300 GPUs, custom 119M-param GPT, val loss 9.5 → 4.08" />
  <project name="Keynoter" note="Tech: Europe Paris AI Hackathon — repo-to-demo-video with cloned voice and AI avatar (Next.js + Remotion + Gradium + fal Aurora)" />
</key_projects>
```

---

*Last updated: May 2026*
*See also: /llms.txt (summary), /mistral.txt (tailored for Mistral applications), /cv (PDF)*