Kaptain
Ongoing research with Prof. M. A. Saied at Université Laval. A Kubernetes scheduler extender that lets a heuristic, a small model or an LLM place pods on a live cluster, and compares each with the default scheduler. Paper in progress.
I'm finishing a dual master's at Université Laval and UTC, and starting as a physical AI intern at Opalin in January.
I like taking things all the way to the point where they run for real, whatever the field.
Ongoing research with Prof. M. A. Saied at Université Laval. A Kubernetes scheduler extender that lets a heuristic, a small model or an LLM place pods on a live cluster, and compares each with the default scheduler. Paper in progress.
Team project. We trained a policy to imitate the Linux scheduler, then let it schedule on its own. In simulation, its mean response time is 10.5% lower than CFS over 60 runs. A hand-written heuristic still does better.
Course project. Can an agent fall into fewer holes while it learns? Over 100,000 training episodes on a narrow, slippery map, a risk-averse agent fell 5,218 times where Q-learning fell 29,191 times. It also stopped reaching the goal.
An Android app and a small server that follow a friend's year-long bike trip to the North Cape. The phone records a point every five minutes, keeps it offline, and sends it when it finds a network. On the road since September. Live map, published with a seven-day delay.
A Jetson Orin Nano with 8 GB of memory. Real-time YOLO11 detection with TensorRT, and a benchmark of six local Qwen models. The one I kept generates 16.8 tokens per second.
An A3 engineering poster of the SO-101 robot arm, generated from its official CAD model.
Turns a 45-minute Theo video into a 2-minute one by cutting only. Claude picks the cuts, ffmpeg makes them.
Guess the city from a random photo. 500 photos, five cities, one leaderboard. Code.
Tap an RFID card on the reader and Spotify switches to that song.
Physical AI intern, Québec City. January to April.
Software engineering intern in Montréal, on Jami, an open-source communication platform. I put a real-time segmentation model into its live video pipeline.