Alan Blanchet
Agents & LLMOps
- LiveKit Agents
- LangChain / LangGraph
- OpenAI Agents SDK
- MCP (builds servers)
- tool-calling
- LLM-as-judge
- Langfuse
- LiteLLM
- Ollama
- LM Studio
- OpenTelemetry
- sentence-transformers
- FAISS / usearch
- RAG
- AGENTS.md / CLAUDE.md
AI Engineer — Agents, LLM Systems & ML
Alan
Blanchet
I build AI for production — agents, LLMs, voice and vision — and I rebuild the models from scratch when I want to really understand them.
- 5+ years in AI / ML
- ~€900k+ across delivered projects
- Builds & runs the team's AI-agent stack
Grenoble, France
R&D Engineer · Neovision
Grenoble · 25 yo
The stack
PyTorch Hugging Face TensorFlow Keras scikit-learn OpenAI Anthropic Ollama LangChain ONNX OpenCV Gradio Jupyter CUDA NumPy pandas Plotly Python Rust TypeScript C++ React Next.js FastAPI Docker Linux PostgreSQL Redis Git GitHub
01 / About
About
I learn by rebuilding things from first principles.
I'm a largely self-taught AI/ML engineer with 5+ years across production computer-vision, speech and LLM systems, now centred on AI agents and LLM apps. I like understanding the layers beneath what I use, so I've built a few of them myself — my own deep-learning framework with a hand-written autograd engine, a real-time voice pipeline, a multi-user GPU-cluster scheduler, a published computer-use agent extension, a prompt-engineering compiler — because I learn by rebuilding things from first principles. I care about clean, well-generalized, strongly-typed code — built with agents, but not vibe-coded — kept honest with a test-driven, edge-case-first habit.
Largely self-taught · 5+ years
Grenoble, France
French (native) · English (bilingual — 3 years schooled in the UK)
Co-author, SPIE paper (2025)
What I work on
- AI agents
- LLM apps
- Voice
- Vision
- Inference
02 / Skills
Skills
The stack, by domain — what I actually build with. · 163 technologies
LLMs
- Claude / GPT / Gemini (API)
- Qwen
- Llama
- DeepSeek
- GPT-OSS
- Mistral
- RAG
- fine-tuning / LoRA
- MoE
- KV-cache
- quantization
- vLLM
- prompt caching
- Batch API cost optimization
Speech — STT
- Whisper / faster-whisper
- NVIDIA Parakeet
- Kyutai
- Silero
- AssemblyAI
- Deepgram
Speech — TTS
- ElevenLabs
- Cartesia
- Deepgram Aura
- Kokoro
- Piper
- Voxtral
- Kyutai / Moshi
Speech — Audio enhancement & VAD
- Silero VAD
- EBEN
- HiFi-GAN / HiFi++
- MP-SENet
- xLSTM-SENet
- VoiceFilter
- DTLN
- RNNoise
- SpeechBrain
- PESQ / STOI / SI-SDR
- SIP telephony (Twilio / Telnyx)
Vision
- RT-DETR / RT-DETRv2
- D-FINE
- YOLO (v8 / NAS)
- DETR
- Faster R-CNN
- SSD
- SAM
- SegFormer
- U-Net
- DINOv2 / v3
- iBOT
- SwAV
- Barlow Twins
- BYOL
- SimCLR
- EquiMod
- ViT / Swin
- MLP-Mixer
- CLIP / OpenCLIP / SigLIP
- Mask R-CNN
- NMS · Hungarian matching · GIoU · TIDE
Image matching & registration
- GIM
- RoMa
- DKM
- LoFTR
- LightGlue
- SuperPoint
- OmniGlue
- RANSAC / MAGSAC
- homography / TPS
Inference & optimization
- TensorRT
- ONNX / onnxruntime / graph surgeon
- INT8 / FP16 / FP8 quantization
- CUDA / Triton
- CuPy
- batching
- einops
Reinforcement learning
- REINFORCE / VPG
- DQN family (Double, Dueling, Recurrent)
- PPO
- R2D2 / NGU-style
- RND
- PER (SumTree)
- Gymnasium
- Ray RLlib
HPC / distributed
- SLURM
- submitit
- Lightning DDP
- Accelerate
- DeepSpeed
- Ray
- Jean Zay (IDRIS)
Infra & Deployment (MLOps)
- Linux (deep) / systemd
- Docker / Compose
- Kubernetes / Helm
- Nginx / Caddy / Traefik + Let's Encrypt
- AWS (S3/SQS/Cognito/CloudWatch/EC2)
- Scaleway (GPU, dedicated endpoints)
- GCP
- Terraform
- Tailscale
- PostgreSQL / Prisma / SQLAlchemy / SQLite
- MariaDB/MySQL
- Redis
- CI/CD
- CVAT
- Roboflow
- FiftyOne
- DVC
- MLflow
- W&B
Languages & Tools
- Python (expert)
- Rust (production)
- TypeScript / JS
- C / C++
- CUDA
- Dart
- SQL
- PHP
- Shell
- PyTorch
- Lightning
- timm
- transformers / datasets
- einops
- torchmetrics
- Albumentations
- sklearn / XGBoost
- polars / pandas
- NumPy
- React
- Next.js
- Svelte
- Vue
- Tailwind
- Flutter (+ flutter_rust_bridge FFI)
- FastAPI
- Express
- Gradio
- Streamlit
- uv / Poetry / conda
- cargo / clippy / miri / bindgen
- just
Soft / domain
- Bilingual FR/EN
- scientific writing & client reporting
- mentoring
- cross-team enablement (R&D News)
- GDPR / sovereignty-aware architecture
- reads & reproduces papers
- code-as-craft convictions
03 / Experience
Experience
R&D Engineer
Neovision · Grenoble · 2022–presentAI / ML at the core, with regular cross-functional support — networking & infra, full-stack, account administration.
~€900k+ delivered ~30-host fleet
- AI enablement — build the in-house agent tooling (MCP servers, multi-provider routing, on-prem LLM serving); run the bi-weekly "R&D News" briefing for the team.
- Run the AI stack — hold the company's API access to the AI providers, manage the AI-agent accounts and budget, and the team's Copilot + Claude accounts.
- Delivered work — lead or primary engineer across ~€900k+ of projects; partially automated the company's R&D tax-credit (CIR) report generation (2025, reused 2026).
- Infrastructure & networking — built and operate a multi-user GPU compute platform from scratch; I run a ~30-host fleet across Scaleway, AWS and on-prem — a dedicated L40S GPU box, a shared 2× RTX 3090 workstation I administer, CVAT, NAS and edge devices — with reverse proxies, TLS and container orchestration.
- Full-stack delivery — React / Next.js / Svelte front-ends through FastAPI / Django back-ends.
How much AI?
Filter by how much was built with my AI agents — from hand-crafted to fully orchestrated.
hand-built agent-built
● built with my orchestrated AI agents ○ hand-written
-
Conversational voice-AI agent — real-time telephony
2025–present Lead tech €300k+- Real-time phone agent, medical appointment scheduling
- pipeline built from scratch → migrated to LiveKit (with the team)
- per-call LLM/STT/TTS with fallbacks
5 more
- tool-calling
- LLM safety check
- full observability
- in production
- in-house demo of the full range, incl. live voice cloning on a call (with explicit consent / GDPR warnings).
- LiveKit
- LangGraph
- multi-provider STT/TTS
- SIP telephony
- Langfuse
- OpenTelemetry
- eval harness
-
Internal AI enablement & agent tooling
2024–present- Help teammates whenever they want to test or experiment with an AI via API
- in-house tooling — MCP servers
- multi-provider routing
3 more
- self-hosted LM Studio server (2× RTX 3090, on-prem/confidential)
- a prompt/standards system that governs how agents write code
- partially automated the company's R&D tax-credit (CIR) report generation (2025, reused 2026).
- MCP servers
- local LLM server (on-prem)
- Ollama / LM Studio
- AI enablement
- coding standards as prompts
-
Multi-head image-classification system
2026 Lead tech €65k- Multi-head image classification (DINOv2 + VLM), delivered for a client
- what mattered most: the home-built agentic system used to build it.
- DINOv2
- multi-head classification
- agent-built
- FastAPI / Gradio
-
Speech enhancement & reconstruction framework
2025 Lead tech €85k- Rebuilds clean speech from in-ear microphones
- several enhancement models (EBEN · HiFi-GAN/++ · MP-SENet · xLSTM-SENet · VoiceFilter) behind one pipeline
- mouth-to-ear acoustic-transfer data simulator (3rd-octave filterbank, per-band gains from measured transfer functions over 28 users)
2 more
- PESQ/STOI/SI-SDR
- shipped (FastAPI · Gradio · Docker · Traefik HTTPS · HF Space). ~36k LOC.
- PyTorch Lightning
- torchaudio
- GANs
- audio denoising
- FastAPI / Gradio
-
Image matching & registration
2025 Lead tech ~€80k- Image registration / template matching for product authentication
- keypoint matching (GIM-DKM · RoMa · LoFTR · LightGlue · SuperPoint · OmniGlue) → RANSAC/MAGSAC → homography/TPS warping
- robust to smartphone capture (lighting · blur · perspective)
4 more
- custom CVAT keypoint-matching annotation plugin
- GTE/CTE accuracy metrics
- automatic keypoint-track propagation into a CVAT pipeline
- Gradio app, packaging, deployment.
- keypoint matching
- RANSAC / homography
- CVAT plugin
- GTE / CTE
- PyTorch
-
Internal ML-demonstrator platform
2024–26 Lead tech- Hosts the company's client demonstrators + internal solutions
- 25+ apps (classification · detection · forecasting · recommendation · photogrammetry)
- each deployed/managed from Docker images, shareable links
3 more
- per-demo orchestration
- reverse proxy
- role-based access. Next.js 16 + Prisma/PostgreSQL + Scaleway S3, self-deploying with Docker + Nginx/Caddy + Let's Encrypt.
- Next.js
- PostgreSQL / Prisma
- Docker orchestration
- Scaleway
-
Real-time object detection
2024- In-depth study + ablation of SOTA detectors (RT-DETR · YOLO-NAS · YOLOv8) → selected RT-DETR (then SOTA)
- TIDE / dedup / INT8 PTQ / MLflow / DVC
- reimplemented RT-DETR on HF Transformers (Pydantic-typed config), grayscale/industrial adaptation
1 more
- edge deployment (ONNX batched_nms · TensorRT INT8/FP16/TF32/FP8 via pycuda).
- RT-DETR
- TensorRT / ONNX
- quantization
- MLflow / DVC
-
DATAWISE — self-supervised-learning benchmark
2024–25 Lead tech ~€400k- DATAWISE — a benchmark of modern self-supervised vision models (DINOv2 · DINO · iBOT · SwAV · Barlow Twins) across ImageNet-1K/22K, CIFAR, Food-101, SUN397, COCO
- clean modular backbone/neck/head
- trained on the Jean Zay national supercomputer (SLURM, Hydra-Submitit sweeps, Lightning DDP, V100)
1 more
- MLflow / W&B / TensorBoard. ~23k LOC. Also reproduced EquiMod (BYOL + LARS) on Jean Zay. A ~€400k programme funded by the Région Auvergne-Rhône-Alpes — currently unfinished and on hold.
- SSL
- PyTorch Lightning (DDP)
- SLURM
- Jean Zay
-
Image + text search platform
2022–24 €500k- Lead front-end (~80%), €500k production platform
- visual pattern extraction (Mask R-CNN)
- image-similarity + free-text search (CLIP/OpenCLIP)
1 more
- also built parts of the embedding-search backend (Django / AWS-SQS / S3). React 18 + TS, AWS Cognito, i18n, crop/zoom annotation UI.
- React / TypeScript
- Mask R-CNN
- CLIP / OpenCLIP
- AWS
- semantic search
-
Clinical bacteria-classification R&D
2023–24- Co-lead ML
- segmentation (U-Net / SegFormer / LR-ASPP) + classification (timm / Transformer) of bacterial colonies vs classical baselines (SVM · stratified K-fold)
- ~40-page scientific report
1 more
- co-author on the peer-reviewed SPIE paper.
- segmentation
- SVM baselines
- cross-validation
- scientific writing
~200
"R&D News" model watch, briefed to the whole team — a short briefing every 2 weeks on notable new models, architectures & techniques. Close follow of SOTA (LMArena · ARC-AGI · SWE-bench…).
04 / Projects
Projects
Current — AI / ML & systems
interact — vision-grounded computer-use MCP
MCP server letting any agent act on what it sees across browser and real desktop (navigate/click/type/scroll/drag); returns text diffs of what changed instead of raw screenshots. GUI grounding fuses VLM detection + the AT-SPI accessibility tree; LiteLLM multi-provider router with cost-aware auto model-selection ranked from public benchmarks (MMMU, ScreenSpot-Pro, Video-MME); isolated software-GL sandbox so GPU/Flutter/Electron apps render. Installs into the major agent clients; files GitHub issues automatically. MIT.
- MCP
- computer-use
- VLM
- LiteLLM
- Rust
- MIT
repo View project
From-scratch DL framework (AI-4-Alan)
My from-scratch deep-learning framework — where I rebuild things from first principles to actually understand them. Its spine is one generic data interface: classification, detection and reinforcement learning all flow through a single Dataset abstraction, pushed so far that a Gym RL environment is itself wrapped as a Dataset — the same dataloader and training loop then drive supervised and RL runs alike. Built on myconf, a hand-written type-driven config system (metaclass coercion, lazy computed fields, no Pydantic), with tensor-subclass modalities (image/text/bbox) and hand-reimplemented ResNet · VGG · ViT · DETR (Hungarian matching) plus the DQN family. Ships a small scalar autograd with graph visualization, written to relearn backprop. ~11k LOC — openly a learning project. Below: an Atari Breakout rollout — the Gym RL environment wrapped as a Dataset — plus the 84×84 grayscale observation the pipeline feeds the agent.
- PyTorch
- myconf
- autograd
- RL
- from-scratch
repo View project
Multi-agent prompt & standards system
The project I'm most attached to. A custom compiler turns 32 declarative "paradigms" (~1,400 lines of engineering principles) + a manifest into conditionally-loaded skills, invocable subagents and per-tool system prompts (Claude Code, Codex, Cursor, Copilot). Encodes a real multi-agent workflow (independent visual-critic gate, skeptical tester, frustration-analyzer, librarian-as-sole-editor). Prompt engineering as compiled, versioned software.
- prompt-engineering
- compiler
- multi-agent
- Claude Code
private View project
any-compute — Rust compute & viz engine
Framework-agnostic compute/visualization engine: SIMD/CUDA/ROCm/MKL/Metal kernels, WGSL/GLSL/SPIR-V shaders, virtualized rendering, C-ABI FFI to Python/JS/WASM. Built almost entirely with my AI agents.
- Rust
- CUDA / ROCm / Metal
- SIMD
- WASM
- FFI
repo View project
Cross-platform learning app (aino)
Flutter + Rust (flutter_rust_bridge) learning app with an LLM content pipeline — local-embedding semantic dedup (fastembed + usearch), batched/cached inference, macro CRUD codegen. Actively maintained.
- Flutter
- Rust
- flutter_rust_bridge
- embeddings
private View project
Recruitment-triage agent
Parses CV/cover-letter PDFs (docling), classifies candidates with an LLM, files them to Google Drive and notifies the team on Slack — scheduled (cron, keyring creds). The kind of agent that may well be reading this CV.
- LLM
- docling
- scheduled agent
- Slack
internal View project
torch-module-observer
A small open-source PyTorch utility that taps any nn.Module with forward hooks to pull out intermediate activations and feature maps — point it at a layer, run an input, get that layer's response back to inspect or visualise. Below: a ResNet's shallow-layer feature maps (the dog still legible per filter) beside a deep layer's abstract activations.
- PyTorch
- hooks
- feature-maps
- open-source
repo View project
desktop-mcp
An MCP server that captures X11 windows and runs vision analysis over them, exposing desktop screen-grounding to any agent — a focused companion to interact for seeing what's on screen.
- MCP
- X11
- VLM
- computer-use
repo View project
linux-commands
My published open-source Linux alias framework (since 2022) — self-documenting per-tool modules (git/apt/npm/django/nvidia/tensorboard/disk-mem/stats), incl. an interactive gum-driven multi-branch auto-rebase with conflict-guard + auto force-push; keyring-backed secret handling.
- Shell
- open-source
- DX
- keyring
repo View project
Earlier work — where it started
Open the archive Games, student builds and first projects — 13 more, kept in the archive rather than shown here.
Just for fun
My Slack avatar runs on my Claude usage
For fun, I wired my Claude usage to my Slack profile picture — so the team can tell at a glance when I'm about to run out (and when I'll be sad about it). A GPT image model paints the mascot through ten moods, Pillow stamps the exact usage bar on top (an image model can't draw precise widths), and a systemd timer reads my session usage from ~/.claude and pushes the matching frame to Slack via users.setPhoto.
Plenty of Claude left Out of usage
A personal user token, scoped so it can only change my own avatar.
05 / Research & teaching
Research & teaching
Publication
Co-author (2nd of 6) on a peer-reviewed SPIE paper — CNN / SVM classification on biomedical multispectral imaging. A low-cost device for label-free, species-level identification of uropathogens; shows 240 spectral channels can be cut to <10 with limited loss.
Leroux, D., Blanchet, A., Davenas, C., Lac, L., Le Bihan, Y., & Fulchiron, C. (2025). A frugal multispectral imaging solution to identify uropathogens via SVM and ANN classification. Translational Biophotonics: Diagnostics and Therapeutics IV, Proc. SPIE Vol. 13934, 1393430.
DOI 2nd author of 6 · ECBO 2025, Munich · published 18 Dec 2025
Mentoring & teaching
Supervised a Master's research internship on LLM-based agents (defended, Grenoble INP/UGA) — a B2B company-data extraction & validation agent; benchmarked LLMs × MCP search tools (Tavily / Exa / DuckDuckGo; SIRET/SIREN validation against official French registries). Coached 5 engineering-student teams at ESISAR (Grenoble INP) on industrial AI projects — incl. a real-time speech-transcription tool needing domain-vocabulary performance.
- internship supervision
- student coaching
- LLM agents
- MCP search
- ASR / domain vocabulary
06 / Education
Education & certifications
-
ML Engineer (work-study)
OpenClassrooms · Neovision
2022–24
-
AI-oriented robotics (work-study)
IMERIR
2021–22
-
DUT Computer Science (3rd/19 · Code Game Jam 2020)
IUT Montpellier-Sète
2019–21
-
Baccalauréat S, mention Bien
2019
-
3 years schooled in the UK — bilingual FR/EN
2010–13
The online ML degree was deliberate — local engineering schools weren't AI-focused enough; the online route freed time to experiment widely and level up faster.
Video archive
07 / What I care about
What I care about
I'd like Europe to keep credible AI options of its own — so I build on European sovereign infrastructure where it holds up, while staying pragmatic: validate with the best models first, then move sovereign once it proves out.
In practice
Self-host open models (Qwen, Gemma) on-prem for confidential data
Preferred sovereign stack — self-hosted, Mistral, Scaleway, Kyutai
08 / Contact
Let's talk
Happy to talk — about AI, a project, or anything you saw here. The fastest ways to reach me: