Section 02 — AI & Machine Learning

AI &
Machine
Learning.

PyTorch 2.10 · MLflow 3.10 Slurm GPU · Linux ───────────── DESS IA Python 3.12 · Docker
Available now

I approach ML with structure — clean pipelines, tracked experiments, reproducible results. The DESS gave me the theory. The real projects gave me the instincts.

PythonPyTorchTransformers (HF)MLflowScikit-learnComputer VisionNLPLinux / SlurmDockerCI/CD
IA
Diploma
DESS — Intelligence Artificielle
Université Laval · 2025 ✓
ML OPS
Certification
MLOps Concepts
DataCamp · Verified ✓
02
Computer Vision · U. Laval · HDRNet
Depth-Guided Retouching
Adapted HDRNet for depth-guided color grading. The depth map tells the model where to adjust and where to leave alone. Trained on MIT-Adobe FiveK, under 2s per image on CPU.
PyTorchHDRNetDepth MapsMIT-Adobe FiveK
<2s
per image
Model
HDRNet · depth-guided
Framework
PyTorch · OpenCV
Dataset
MIT-Adobe FiveK
Perf.
<2s / image · CPU
03
MLOps · Open Source · Ongoing
OFF Canada Contributor
Started by exploring all three scoring models: Nutri-Score, Nova Score, Green Score. Then moved into direct contribution — building the Nutri-Score pipeline, leading a student team on the other two. Clean data and tracked experiments are what the next model runs on.
TransformersMLflowOpen SourceTeam Lead
3
models
Stack
Transformers HF
Tracking
MLflow · versioned
Dataset
Open Food Facts CA
Status
Ongoing · 3 models
04
ML · Hardware · Signal Processing · U. Laval
Sleep Quality Scoring
Sleep quality scoring from physiological signals: ECG, respiration, movement, temperature. The model predicted a sleep efficiency score that matched what the test subject gave their own night. K-means for phase detection, Random Forest for the final score. Around 80% accuracy.
Random ForestK-meansSignal ProcessingHardware
80%
accuracy
Signals
ECG · Respiration · Temp.
Models
K-means · Random Forest
Task
Sleep efficiency score
Accuracy
~80%
05
NLP · Computer Vision · Research · U. Laval
SignTerpreter
Research on ASL-to-English translation — not word-by-word, not word concatenation, but full coherent sentences. Pose sequences only, no RGB, no intermediate glosses. We adapted Uni-Sign into a lightweight modular architecture: pose-only pipeline, x5–x10 memory reduction, single-stage training on OpenASL. BLEU-4 stayed low. Generating coherent sentences from gestures alone is genuinely hard. But the architecture is clean, reproducible, and the text backbone is swappable.
Uni-SignPose EstimationNLPOpenASL
x5 – x10
memory reduction
Input
Pose sequences · no RGB
Base
Uni-Sign · adapted
Dataset
OpenASL + How2Sign
Status
Research · U. Laval
06
Core ML · iOS · In progress
AI Music Practice App
On-device ML model for a music practice app. Note prediction and harmonic guidance, running entirely on iPhone/iPad. In development. A musician plays something, the model responds — no server, no wait.
Core MLSwiftiOSOn-device
0 ms
server latency
Runtime
Core ML · on-device
Task
Note prediction · harmony
Platform
iOS
Status
In progress
2M+
Data points processed
6+
ML projects