What I build and run on my own time.
Two second-hand Lenovo mini PCs in France, bought for €200 together, run my services on Docker Compose. I drive them over SSH from Sydney, 17,000 km away; they back themselves up and come back on their own after a restart.
I record voice notes on my iPhone in Sydney. A mini PC in France transcribes them with Whisper, then an LLM files each one into my Obsidian vault as linked notes, an idea from Andrej Karpathy, and the vault syncs back to my phone and Mac.
I write software and run my Linux machines over SSH with coding agents (Hermes Agent, Codex, Claude Code) and MCP-connected tools. I've also connected self-hosted inference endpoints to agent workflows.
In 2026 I served open-weight models on a Lenovo ThinkStation PGX, Lenovo's take on the NVIDIA DGX Spark: Qwen3.8-27B, Qwen3.5-122B, Nemotron 3 Super, Gemma 4 and DeepSeek V4 Flash, with vLLM, SGLang and llama.cpp. My first Hermes Agent setup ran on it.
A 2-bit DeepSeek V4 Flash, about 97 GB of weights in 128 GB of memory, read a 420,000-token prompt in a single session on an experimental runtime. Qwen3.8-27B took 260,000-token prompts and reached 32 tokens a second with speculative decoding. On a tool-use benchmark of 84 scenarios, the best of 14 configurations scored 92/100.
The task was to estimate biomass from images. Starting from existing work, I ran the DINOv3 fine-tuning and the submissions, including staged unfreezing of the backbone.
Work and studies, newest first. The full version is on the CV.
2025-04 – 2026-07France
AI Technical Project Engineer
2024-04 – 2024-09Tarnos, France
Data Scientist, Health Monitoring (Intern)
2022-04 – 2022-08Vernon, France
Computer Vision / R&D Intern
2017 – 2024France · Spain
Centrale Nantes · Diplôme d'Ingénieur (Master's-level degree), General Engineering · 2020–2024
Universidad Politécnica de Madrid · exchange year, machine learning and data science coursework · 2022–2023
Université de Bordeaux · Bachelor's Degree in Physics · 2017–2020
languages
English: professional working proficiency · French: native · Spanish: intermediate
Questions people ask me
At Safran Helicopter Engines I built a Python regression model that estimates fuel flow on Arriel 2E and 2D engines from physical engine parameters. I integrated it into Safran's industrial pipeline with unit tests and shipped it through the team's existing GitLab CI/CD to production. A year later it was still used every day to monitor every Arriel 2E the team tracked.
Track record · SafranA forward-deployed engineering role in Sydney, between the technical work and the client: building with LLMs and agents, and staying close to the people who will use them. I have worked on both sides: a model taken to production at Safran, and a year on the customer side of an NLP project at the French Defence Procurement Agency.
CVI live in Sydney and hold a Working Holiday visa, subclass 417.
What “done” really means. I wrote what a contractor had to build, then sat with the users to check what came back, and most of the work lives in the gap between a specification and what people actually need. It also taught me to take responsibility, and to make myself heard and hold my ground, often as by far the youngest person at the table. Now I want to be on the other side of that table, building.
Track record · DGAOr gaetan@forestier.dev · LinkedIn · GitHub (Taelyos) · CV · PDF