Tolérance aux pannes et mise à l'échelle Kubernetes
Un service Python conteneurisé sous HPA Kubernetes, soumis à des tests de charge pour vérifier qu'il s'adapte et se répare seul — 99 % de disponibilité, mesurée par Prometheus.
Les travaux d'infrastructure d'abord. Les projets data et machine learning plus anciens sont regroupés à part.
Un service Python conteneurisé sous HPA Kubernetes, soumis à des tests de charge pour vérifier qu'il s'adapte et se répare seul — 99 % de disponibilité, mesurée par Prometheus.
A Terraform-provisioned AWS stack that absorbed a 20,000-agent denial-of-service attack while holding CPU below 50%, with a dashboard reading CloudWatch metrics live.
A convolutional autoencoder whose decoder is trained to strip noise out of its input rather than just reconstruct it. Built during the Ericsson internship.
Snake rebuilt from scratch so every rule was controllable, then NEAT agents that evolve their own neural networks to play it.
A hackathon tool tracking facial landmarks and head movement to estimate whether someone is focused — and a decent argument for why that should stay a hackathon tool.
Semi-empirical wildfire spread modelled over real terrain reconstructed from satellite imagery in Unity, accounting for elevation, wind and vegetation. Published on ResearchGate.
A collaborative effort to scrape scattered Algerian historical, social and economic data and turn it into visualisations that actually explain something.