# Edouard Foussier — Applied AI Engineer (tailored for Mistral · Proto Team)

<candidate_profile>
  <name>Edouard Foussier</name>
  <role>Applied AI Engineer — Full-stack systems · Prototyping</role>
  <location>Paris, France (available immediately, open to relocation)</location>
  <target_role>Applied AI Engineer, Prototyping — Mistral AI Proto Team</target_role>
  <specialty>Customer-facing AI prototypes, full-stack delivery, RAG, multimodal vision, agentic systems</specialty>

  <tagline>
    I ship AI prototypes at hackathon speed and harden them for production.
    Three podiums in 2026, two paying SMB clients shipped on modern full-stack TypeScript / Svelte
    stacks with multimodal AI, two Mistral hackathons (one product still live in production),
    one sovereign RAG in beta for the French State — same person, same approach: build the working
    thing, measure, iterate.
  </tagline>
</candidate_profile>

---

## Why Proto Team

<fit>
  <signal>Prototype-over-proposal builder: won the beta.gouv government mission by shipping a working demo instead of a slide deck. Same instinct drives the hackathon wins and the multi-phase SMB rollouts (Phase 0 shipped in days to de-risk, Phase 1 on a fresh production stack).</signal>
  <signal>Hackathon-trained for 4-to-8-week intensity: 3 podiums in 2026 (Voodoo &amp; Anthropic 1st, GOSIM/Z.AI 1st, GPU Mode Paris 3rd) plus a Mistral hackathon product still live in production.</signal>
  <signal>Two Mistral hackathons in 2026: Le Sphinx at the Mistral AI Worldwide Hackathon (live at thesphinx.ai), and the Alan × Mistral AI Health Hack (April 2026). Already wired into the Mistral ecosystem.</signal>
  <signal>Full-stack range: FastAPI + Pydantic backends, Next.js 16 / React 19 / Svelte 5 frontends, PostgreSQL + Prisma, NextAuth, Docker, Caddy, Tailscale, GCP. Two paying SMB clients running on this stack today.</signal>
  <signal>Multimodal AI in real production use: Claude Vision for vision-based supplier delivery verification (Déjà Bu); voice agents with ElevenLabs and Mistral Large on Le Sphinx; Z.AI vision + text on Xiexie via a Cloudflare Worker translation layer.</signal>
  <signal>Operational scale-up muscle: ran Le Wagon Paris (largest campus) from ~120 to ~450 students/year and from ~€540K to ~€3.1M ARR while managing technical education. Then closed CAC40 deals at a co-founded digital studio (Vili &amp; Ve).</signal>
</fit>

---

## 2026 Hackathon Track Record

<hackathons>
  <hackathon date="2026-05">
    <event>Voodoo &amp; Anthropic Hackathon</event>
    <track>Track 3 — Market Intelligence</track>
    <result>1st place</result>
    <product>
      VoodRadar — takes a mobile-game name and produces 3–5 ready-to-test ad video concepts in minutes.
      Pipeline: SensorTower data → Gemini 2.5 Pro Vision (creative deconstruction) → Claude Opus
      (brief generation) → Scenario + Kling (video generation) → FFmpeg (audio mix).
    </product>
    <format>30 hours, team of 3.</format>
    <stack>Python 3.12, FastAPI, Pydantic 2, React 19, TanStack Router, shadcn/ui, SSE pipeline, ~125 MB cached knowledge base, 20+ endpoints.</stack>
    <repo>https://github.com/edouardfoussier/voodoo-anthropic-hack-winner</repo>
  </hackathon>

  <hackathon date="2026-05">
    <event>GOSIM Agentic Hackathon</event>
    <track>Z.AI Innovation Track</track>
    <result>1st place</result>
    <product>
      Xiexie — voice-first macOS companion that protects seniors from email scams in real time.
      Native Swift app (ScreenCaptureKit, push-to-talk, ember overlay) with a Cloudflare Worker
      bridging Z.AI's vision and text models via OpenAI-compatible payloads.
    </product>
    <format>36 hours, solo. Live at getxiexie.com.</format>
    <stack>Swift, ScreenCaptureKit, Cloudflare Workers, Z.AI GLM-4.6 (text), Z.AI GLM-4.5V (vision), AssemblyAI streaming STT, ElevenLabs TTS, custom ONNX wake-word.</stack>
    <repo>https://github.com/edouardfoussier/gosim-zai-hack-winner</repo>
  </hackathon>

  <hackathon date="2026-04">
    <event>GPU Mode Paris Hackathon (Sesterce / NVIDIA B300)</event>
    <track>LLM Pre-training</track>
    <result>3rd place</result>
    <details>
      One-day LLM pre-training competition on 32 × NVIDIA B300 GPUs, fixed dataset (~49B tokens, 32K vocab),
      10 minutes of GPU per submission. Built a custom 119M-parameter GPT (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. Ran 16 experiments in one day,
      validation loss 9.5 → 4.08.
    </details>
    <repo>https://github.com/edouardfoussier/gpu-mode-hackathon-2026</repo>
  </hackathon>

  <hackathon date="2026-03">
    <event>Mistral AI Worldwide Hackathon (#1 of 2 Mistral hackathons in 2026)</event>
    <product>
      Le Sphinx — voice-powered Akinator-style guessing game built natively on Mistral Large + Small.
      Includes an MCP server for AI agents (mcp.thesphinx.ai), Human-vs-AI duel mode with 4 Bedrock-powered
      opponents, and a theatrical Sphinx powered by Mistral Large + ElevenLabs TTS. Shipped to production
      at thesphinx.ai.
    </product>
    <stack>Mistral Large + Small, FastAPI, Pydantic, ElevenLabs TTS, Web Speech STT, MCP (FastMCP), 220 pytest tests, GCP VM + Caddy + Docker.</stack>
    <repo>https://github.com/edouardfoussier/lesphinx</repo>
  </hackathon>

  <hackathon date="2026-04">
    <event>Alan × Mistral AI Health Hack (#2 of 2 Mistral hackathons in 2026)</event>
    <product>
      Curated 30-builder event hosted by Mistral AI and Alan in Paris — 12 intense hours building AI-native
      health products. Participated as one of the selected builders.
    </product>
  </hackathon>

  <hackathon date="2026-05">
    <event>{Tech: Europe} Paris AI Hackathon — Keynoter</event>
    <product>
      Built Keynoter in 24 hours with a team of 2: turns a GitHub repo into a motion-designed demo video
      with the user's cloned voice and an AI avatar. Contributed voice cloning + TTS pipeline (Gradium),
      avatar lipsync (fal Aurora), multilingual script generation, an experiments lab (3-way TTS comparator,
      4-way photo animation comparator, ~58s Aurora silent-truncation stress test), and a post-hack live Q&amp;A spike with Anam.
    </product>
    <stack>Next.js 16, React 19, Remotion 4, OpenAI Codex Cloud, AI SDK v6, fal AI, Gradium, Postgres + Drizzle, Tailwind v4, Bun, Docker + Traefik.</stack>
    <repo>https://github.com/edouardfoussier/tech-europe-hack-keynoter</repo>
  </hackathon>
</hackathons>

---

## Production Project — Sovereign RAG for the French State

<production_project>
  <client>beta.gouv.fr / DGAFP</client>
  <users>~70 beta HR officers navigating complex legal &amp; administrative documentation.</users>
  <pipeline>
    Multi-stage RAG: intent classifier → parallel retrieval across 4 heterogeneous government sources
    (Légifrance API, service-public.gouv.fr XML, ministerial XLSX + PDFs, internal HR portal) → section
    expansion → reranking → LLM Selector → ContextBuilder with token-budget expansion → cited generation.
    Three architectural iterations (V1 chunked / V2 semantic / V3 hybrid), each benchmarked and redeployed.
  </pipeline>
  <stack>
    Custom RAG (no LangChain/LlamaIndex), PostgreSQL + pgvector, RAGAS + MLflow,
    FastAPI + Pydantic, Mistral, Qwen3 (ablation tests).
  </stack>
  <how_i_won>Delivered a working prototype instead of a proposal.</how_i_won>
  <results>
    <metric>1,490-question goldset, 8,293 evaluation runs across 10+ architecture variants.</metric>
    <metric>RAGAS scores +13% to +37% across faithfulness, relevancy, correctness, context precision and recall (V1 → V3).</metric>
    <metric>+122% answer correctness vs vanilla LLM without RAG (0.30 → 0.67).</metric>
    <metric>74.8% of V3 answers preferred over live beta by GPT-4.1 head-to-head.</metric>
    <metric>Diagnosed retrieval (not generation) as the primary failure mode — reshaped where engineering effort went.</metric>
  </results>
</production_project>

---

## Recent SMB Client Work (2026)

<client_projects>
  <project>
    <name>Archipel — Sales Steering Dashboard for a Kombucha Producer</name>
    <client_url>https://archipelkombucha.com</client_url>
    <pitch>
      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."
    </pitch>
    <stack>
      Next.js 16 (App Router), React 19, TypeScript, Tailwind v4, shadcn/ui, Recharts, Leaflet,
      PostgreSQL (Neon), Prisma v7, NextAuth v5, Sentry, Vitest (74 unit + 20 integration),
      GitHub Actions, Vercel.
    </stack>
    <highlights>
      Multi-source ingestion pipeline (auto-detection, SHA-256 file dedup, atomic transactions).
      Product alias system reaching 88% direct + 100% distributor matching to a 44-SKU catalog.
      Excel forecast import (~3,488 entries, 96% catalog matching).
      Production discipline: multi-user auth, monitoring, 94 tests, CI/CD.
    </highlights>
    <why_it_matters_for_proto>
      Real customer-facing delivery for an SMB, in modern full-stack TypeScript — exactly the kind
      of 4-to-8-week prototyping cycle the Proto Team is built around. Direct relationship with
      the founder/end-user; UX bar set by the customer; production discipline from day one.
    </why_it_matters_for_proto>
  </project>

  <project>
    <name>Déjà Bu — Mobile-First PWA · Inventory + Vision-Based Delivery Verification</name>
    <client_url>https://deja-bu.com</client_url>
    <pitch>
      Private 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.
    </pitch>
    <stack>
      Svelte 5 (runes), TypeScript, Vite, vite-plugin-pwa, html5-qrcode, FastAPI,
      SQLAlchemy 2.x async, Alembic, Pydantic v2, Postgres, Claude Vision, Vercel + Railway.
    </stack>
    <multi_phase_rollout>
      Phase 0 (live): single-file index.html mini-collector shipped in days — camera EAN-13 /
      UPC-A / Code-128 scan, localStorage with 30s heartbeat, CSV export, installable as a PWA
      on iPhone via Safari.
      Phase 1 (in progress): fresh production stack — Svelte 5 + FastAPI + Postgres + Claude Vision.
      Phase 2 (planned): order/delivery reconciliation via n8n.
    </multi_phase_rollout>
    <why_it_matters_for_proto>
      Real multimodal use case (vision-based supplier delivery verification). Multi-phase
      de-risked rollout — ship to validate in days, harden second. Modern frontend stack
      (Svelte 5 with runes). Mobile-first, offline-first, installable. End-to-end ownership.
    </why_it_matters_for_proto>
  </project>
</client_projects>

---

## Technical Range

<skills>
  <category name="Applied AI &amp; Prototyping">
    End-to-end customer-facing prototypes (4–8 week cycles), multi-phase de-risked rollouts,
    customer relationship from scoping to deployment.
  </category>

  <category name="RAG &amp; Retrieval">
    RAG from scratch, hybrid search (BM25 + dense), reranking, multi-source retrieval with routing,
    PostgreSQL + pgvector, structure-aware chunking.
  </category>

  <category name="Multimodal &amp; Agentic">
    Vision APIs (Claude Vision, Gemini Vision, Z.AI GLM-4.5V, OCR, EAN/UPC scanning), voice agents
    (ElevenLabs, AssemblyAI, Whisper), MCP servers (FastMCP, SSE), LiteLLM routing (cloud + local
    models), tool calling, structured outputs (Pydantic), Cloudflare Worker translation layers.
  </category>

  <category name="LLM Training">
    Distributed DDP training on 32 GPUs, Muon optimizer, WSD schedules, RoPE, ReLU²,
    U-Net skip connections, torch.compile, transformer internals.
  </category>

  <category name="Full-Stack Delivery">
    Python (FastAPI, Pydantic, SQLAlchemy 2.x async, Streamlit), TypeScript (Next.js 16,
    React 19, Svelte 5, Vite, Remotion), PostgreSQL + Prisma + Drizzle, NextAuth, Tailwind,
    shadcn/ui, PWA / mobile-first, native Swift (macOS).
  </category>

  <category name="Infrastructure &amp; MLOps">
    Docker, Caddy, Tailscale, Traefik, self-hosted European infrastructure, CI/CD (GitHub Actions,
    Vercel, Railway, Bun), Sentry, MLflow, GCP, Cloudflare Workers.
  </category>

  <category name="LLM Providers">
    Mistral (primary — Large + Small in Le Sphinx prod), Claude, OpenAI (incl. Realtime), Gemini,
    Z.AI GLM, Bedrock, local models (Ollama, Gemma).
  </category>

  <category name="Evaluation &amp; Quality">
    RAGAS (faithfulness, answer relevancy, correctness, context precision, context recall),
    custom goldsets, LLM-as-judge, human-in-the-loop feedback, 8,000+ runs on production RAG.
  </category>
</skills>

---

## Experience

<experience>
  <position current="true">
    <title>AI Engineer — Freelance</title>
    <period>Mar 2026 — present</period>
    <highlights>
      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).
    </highlights>
  </position>

  <position>
    <title>AI Engineer — beta.gouv.fr / French State (DGAFP)</title>
    <period>Sep 2024 — Feb 2026</period>
    <highlights>
      Technical lead for a sovereign RAG assistant for HR officers. ~70 beta users, 8,293 evaluation
      runs across three architecture iterations, 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.
    </highlights>
  </position>

  <position>
    <title>AI Engineer — The Artificial Business</title>
    <period>Nov 2024 — Mar 2025</period>
    <highlights>
      Contributed to a self-hosted AI agent platform focused on sovereignty and GDPR constraints.
    </highlights>
  </position>

  <position>
    <title>General Manager — Le Wagon Paris</title>
    <period>Mar 2016 — Aug 2022</period>
    <highlights>
      Ran Le Wagon's largest campus. Scaled operations from ~120 to ~450 students/year and revenue
      from ~€540K to ~€3.1M ARR while teaching and managing technical education.
      Trained hundreds of developers. NPS 70. Deeply embedded in the French startup ecosystem.
      Returning in 2026 to teach three upcoming Data Science bootcamp courses (Generative AI &amp; RAG,
      Reinforcement Learning, Generative AI for Images).
    </highlights>
  </position>

  <position>
    <title>Co-founder &amp; CEO — Vili &amp; Ve (digital studio)</title>
    <period>Sep 2014 — Dec 2015</period>
    <highlights>
      UX/UI and web/mobile studio. Closed CAC40 clients (Carrefour, EY). Founder-led sales,
      pre-sales, delivery, and project management.
    </highlights>
  </position>
</experience>

---

## Availability

<availability>
  <status>Open to full-time Applied AI roles (also available freelance via Malt.fr).</status>
  <location>Paris, France — open to relocation.</location>
  <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_roles>
  <target_companies>Frontier-model labs, applied-AI startups, customer-facing prototype teams.</target_companies>
</availability>

---

## Contact

<contact>
  <cv>https://edouardfoussier.com/cv</cv>
  <website>https://edouardfoussier.com</website>
  <full_profile>https://edouardfoussier.com/llms-full.txt</full_profile>
  <github>https://github.com/edouardfoussier</github>
  <huggingface>https://huggingface.co/edouardfoussier</huggingface>
  <linkedin>https://www.linkedin.com/in/edouardf/</linkedin>
  <email>edouardfoussier@gmail.com</email>
  <project_sphinx>https://thesphinx.ai</project_sphinx>
  <project_xiexie>https://getxiexie.com</project_xiexie>
  <project_archipel>https://archipelkombucha.com</project_archipel>
  <project_dejabu>https://deja-bu.com</project_dejabu>
</contact>

---

*Last updated: May 2026 — tailored for the Applied AI Engineer (Prototyping) role at Mistral AI.*
*See also: /llms.txt (summary), /llms-full.txt (full profile), /cv (PDF).*