Denoising Diffusion Probabilistic Models (DDPM):
From-scratch implementation in PyTorch of the seminal paper by Ho et al. (2020). The model learns to reverse a gradual noising process over T=1000 timesteps, generating new images from pure Gaussian noise. The architecture is a U-Net conditioned on the timestep via sinusoidal embeddings, with ResNet blocks, multi-head self-attention, and skip connections. Training uses an EMA (Exponential Moving Average) of the model weights for improved sample quality. Trained and evaluated on MNIST and CIFAR-10. Realized in collaboration with CHABANE Oualid. The complete source code is available on my GitHub.

English-to-French Machine Translation (Europarl):
Comparative study of three translation approaches of increasing complexity, trained and evaluated on the Europarl v8 parallel corpus (OPUS). The three systems are: a word-for-word baseline using a bilingual positional dictionary, a cross-lingual embedding retrieval system using multilingual sentence embeddings, and a fine-tuned Seq2Seq Transformer (MarianMT / Helsinki-NLP opus-mt-en-fr) with hyperparameters selected via grid search. The fine-tuned transformer significantly outperformed both baselines (BLEU 0.32 vs. 0.05 and 0.04). Analysis includes attention map visualization, n-gram BLEU breakdown, and sentence-length impact. Realized in collaboration with Maxime Hayakawa and Juan Kenichi Sutan. The complete source code is available on my GitHub.

Scientific Article Information Retrieval:
Built a full information retrieval pipeline for a scientific citation prediction challenge on a corpus of 20,000 research papers. Starting from sparse baselines (TF-IDF, BM25), we progressively moved to dense retrieval using large pre-trained encoders (BGE-large, GTE-large, Granite R2), combining them through Reciprocal Rank Fusion (RRF) with grid-searched weights. A key contribution of our approach is a hand-crafted ontological domain bonus that boosts candidate papers sharing the same scientific domain as the query, which proved to be the single most impactful component of the pipeline. We also explored body-level chunk embeddings with MaxSim scoring, citation context embeddings, sparse signal fusion, and a Matryoshka + LLM re-ranking pipeline using Llama 3.2 via Ollama. Realized in collaboration with Lubin LONGUEPEE, Oualid CHABANE and Lounes KEBDI. The complete source code is available on my GitHub.

IoT Edge: Sensor Monitoring with MQTT, Prometheus & Grafana:
Full IoT edge computing project built on a Raspberry Pi, combining hardware sensing, real-time messaging, and containerized observability. An HC-SR04 ultrasonic sensor publishes distance measurements every 0.5s via MQTT (Mosquitto), which are then bridged to Prometheus through a custom mqtt-exporter and visualized in Grafana dashboards: a proximity indicator (color gauge) and a derived speed metric. The entire stack runs as 5 Docker containers orchestrated with docker-compose, deployable identically on any machine. The project also includes email alerting, a QoS analysis across all three MQTT levels, a sound feedback system that maps distance to musical notes in real time, and a basketball detector designed to run across multiple Raspberry Pis simultaneously with a central score-counting referee script. Realized in collaboration with Isabel FABREGA, Johan MARTIN-BORRET and Zineddine KERMADJ. The complete source code is available on my GitHub.

Ant Colony Rush:
A strategic simulation game built in Java Swing where the player manages an ant colony threatened by a hungry frog. The objective is to collect resources, expand the nest by recruiting new ants, manage their energy and health, and accumulate enough supplies to migrate the colony to a safe territory before the frog devours them all. The frog has a dynamic field of vision and hunger level, forcing the player to plan efficient collection routes while avoiding the predator. The game features multiple difficulty levels, animated menus with background music, a day/night cycle, weather effects, and a scoring system. Realized in collaboration with Johan MARTIN-BORRET, Ania GAROUI, Esther DEBA. The complete source code is available on GitHub.