Hi, I'm Benoit Desnos
AI/LLM Architect with 20 years of engineering. I build production systems from scratch -AI pipelines, identity protocols, developer tools. Alone if needed, then with a team.
AI / LLM Systems
- YokeBrain (private) - AI-powered business model generator. 6-phase pipeline orchestrating ~450 LLM calls via bisociation theory (Koestler). Embeddings, multi-criteria reranking, ontology-driven generation, cognitive bias integration (175+ biases across 7 roles). NestJS, Vue 3, PostgreSQL, Kubernetes.
- melchizedek
- MCP Server for Claude Code. Hybrid search engine combining BM25, vector similarity, and reranking over conversation history. 100% local, zero API keys. Published on npm.
- arXiv Scanner (private) - Pipeline ingesting ~5K arXiv papers, LLM-powered triage (Haiku for fast filtering, Sonnet for commercial scoring across 6 dimensions), identifies papers transformable into products. TypeScript, SQLite, Claude API.
- Kalamos (private) - Local-first LinkedIn editorial manager. MCP Server with 10 tools for Claude Code: hybrid search (BM25 + vectors + RRF + reranker), full lifecycle (draft → schedule → publish → analytics), OAuth2 publishing, anti-detection corpus scraping via Patchright. TypeScript, SQLite, sqlite-vec, ~9K LOC.
- LLM Agent Pipeline (private) - LinkedIn post generator using 9 parallel LLM agents. Each agent evaluates tension/angle pairs on content drafts via structured output (Zod schemas), automated congruence scoring, voice profile extraction. TypeScript, Claude API.
- CE Differentiator (private) - CNN image classifier distinguishing authentic EU CE marking from counterfeit Chinese export marking. 151K synthetic training images (Albumentations, Perlin noise). TensorFlow/Keras, OpenCV.
- lego-emcom - Emergent communication framework for 3D structure reconstruction. Two neural agents learn to communicate about colored LEGO structures through a discrete channel -- no pre-defined language, no supervision on messages. 3D CNN encoder/decoder, Gumbel-Softmax channel. PyTorch. Paper in preparation.
- ml-sonets - Self-generative sonnets using autoencoder (TensorFlow/Keras). Trained on 4,820 classical French poems.
Reverse Engineering & IoT
Cryptography & Identity
- OpenSSL compiled to WebAssembly -70+ elliptic curves, 28 hash algorithms, full crypto in the browser
- ISO 18013-5 (mDL) and ICAO 9303 (MRTD) implementations from scratch in TypeScript
- France Identite government app reverse engineering (Frida, Ghidra)
Tech stack
AI/ML - LLM pipelines (Claude, structured output, Zod), MCP SDK, embeddings, reranking, RAG, TensorFlow/Keras, Transformers.js
Backend - Node.js, NestJS, Kafka, PostgreSQL, YugabyteDB, Redis, SQLite
Infrastructure - Kubernetes, Docker, Helm, CI/CD, Prometheus, Grafana
Security - OpenSSL, PKCS#7, ECDH, PACE/BAC, NFC protocols, BLE reverse engineering
Languages - TypeScript, JavaScript, Python, Swift, C
Frontend - Vue.js
Career highlights
| Period | Role | Scope |
|---|---|---|
| 2024–now | Staff Engineer, SpringCard | Digital identity (mDL, MRTD, EUDI), SaaS B2B (110k LOC), OpenSSL-WASM, sole developer |
| 2021–2024 | Tech Lead, Groupe CREATIVE | Agile projects (200–1000 man-days), 8-person team |
| 2020–2021 | Senior PM, Inetum | SI Vitale program (CNAM), 12-person team within 90 FTE program |
| 2018–2020 | CTO, BeMyApp | Hackathon SaaS (Microsoft, Google, L'Oréal), team 4 to 8, -80% hosting costs |
| 2016–2018 | CTO, iLearn | ML/speech recognition R&D, TensorFlow feasibility study |
| 2015–2016 | CTO, MobilisPro | Two-sided SaaS (film, TV, radio, games), 1st tech hire, team of 5–7 |
| 2012–2015 | PM / Mobile Lead, Netapsys | Mobile projects (iOS/Android), 3–10 person teams |
| 2010–2011 | PM / Analyst, Infotel | Detached at SIGMA (E.Leclerc) |
| 2006–2010 | Functional Designer, Sogeti | Enterprise clients (CNIEG, ENIM, MMA, France Telecom) |
