Francois G.

Co-founder of Leonard Intelligence to offer AI products that run on the client's premises. Open-source models on private cloud or on-premise, with data, security, and compliance remaining with them.

I manage product and engineering. I'm in the room for the first client call, and at the keyboard when the code goes to prod. Small team, no hand-offs.

Aura (product): Agency management SaaS.

CRM, billing, mission tracking, emailing, analytics, and social calendar in a single app. Includes Promethee, an AI assistant with function calling that queries the business database, creates contacts, writes prospecting emails, and tracks its own token consumption. 23 pages, 56 REST endpoints, 20 SQLite tables with automatic migrations.

Stack: Next.js 16, React 19, TypeScript, Node.js 22, SQLite (WAL mode), Tailwind 4, shadcn/ui, Google Gemini (2.5 Flash), MailerLite, Docker.

Applications on Mistral AI: cross-platform (desktop + web).

4 distinct applications sharing the same React codebase, distributed on Electron and in the browser. OCR with Pixtral vision (single and batch mode, 7 document types including invoices, IDs, CVs, shipping labels). A RAG explorer built from scratch: PDF extraction, chunking with overlap, vector search, anti-hallucination safeguards with mandatory source citations. Real-time meeting transcription with a Node.js WebSocket proxy relaying a stereo PCM audio stream (microphone and system audio) to Voxtral at 16 kHz, plus a Mistral Small pass providing live coaching to the speaker during the meeting. Full audio file transcription with speaker diarization and follow-up Q&A on the obtained text.

Stack: React 18, Electron 35, Node.js, Mistral AI (6 endpoints), Web Audio API + AudioWorklet for PCM capture, WebSocket, SSE streaming.

Co-founder of Melies AI and sole developer of the platform for 2 years.

Melies.ai transforms a text prompt into a finished film. Briefing, script, characters, images, voice, video, FFmpeg editing, publishing. ~40,000 lines of TypeScript in production, 87 API endpoints, 38 pages, 101+ business logic files. Backend, frontend, AI integrations, infra, deployment.

What I built:

• Multi-provider LLM routing on OpenAI, Anthropic, Google Vertex, Groq, OpenRouter, Fireworks, Replicate, FAL

• Video generation pipeline calling 6+ models (Vertex Veo 2/3, SORA 2, Seedance) alongside image models (Flux Pro, DALL-E, Gemini), voice and music via ElevenLabs, final editing with FFmpeg

• Background job system on Bull + Redis, 50 jobs in parallel

• Custom Auth: Google OAuth with FedCM, JWT in HTTP-only cookies, refresh token rotation

• Vertical video feed like TikTok with iOS Safari workarounds: video element pool to bypass the 16 decoder limit, audio unlock on first gesture, layouts adapting to the keyboard

• Stripe billing

• Admin dashboard

• Content moderation with Hive as primary and a VLM as fallback for NSFW, violence, hate symbols

• VPS deployment on Docker Compose (6 orchestrated services)

Stack: Next.js 15, React 18, TypeScript, PostgreSQL 15 + Prisma 6, Redis 7 + Bull, LangChain, 10+ AI providers, Stripe, FFmpeg, Framer Motion, Radix UI, SSE, Docker Compose, Cloudflare Tunnel.

MEL token system on Base (L2 Ethereum).

I also designed and wrote the entire token smart contract stack: 4 Solidity contracts. ERC-20 with ERC20Permit. Multi-round ICO with Chainlink oracles and Uniswap V2 swaps. Staking with 5 lock durations (up to an x3 multiplier), auto-compounding, and burning in case of early unlock. ~12,000 lines of Solidity, 343 Foundry tests, ready for audit.

Stack: Solidity 0.8.20, Foundry, OpenZeppelin v5, Chainlink, Uniswap V2, Base L2.

4 years as a Solidity trainer at the largest blockchain school in France. I built and taught the smart contract program to several cohorts of career changers, coming from Web2, DevOps, or product backgrounds. Topics covered: Solidity, contract architecture, DeFi patterns, and dApp design. Many of my former students are now employed in the ecosystem.