Erwan DAVID's repositories
GraphRAG-RAG_LLM_Specialization_Benchmark
This project implements and compares RAG and GraphRAG for updating LLMs with recent events (e.g., 2024 U.S. election). Using the Trump24-25 benchmark with human and automatic evaluations, we show that while web access performs often very well, RAG and GraphRAG are solid options for private or specialized data depending on the structure of the data.
LOTEREO-2D_arcade_platformer_game
This Unity project in C# is a 2D arcade platformer. The player's objective is to pass through the door on the last (2nd) level. To open the doors on each level, which lead to the next level, the player must kill a certain number of AIs on the map.
Natural_evolution_of_an_ecosystem
This multi-month Python project is a simulation of natural evolution, allowing complex behaviors to emerge. Individuals are driven by artificial neural networks and evolve through genetic mutations. They must learn to develop individual and group strategies to survive in this environment.
Rain_gauge_anomaly_detection
The aim of this Python project is to detect anomalies in the rainfall totals of a network of rain gauges by comparing the totals between neighbouring rain gauges and precipitation radars.
Random_procedural_environment
The aim of this Python project is to generate a semi-random map representing a complex, realistic environment seen from above in 2D.
CNN_embedding_matching
This is repository contains our work for an AI-Challenge with HeadMindPartners, taking part of CentraleSupélec 3rd year
CNN_number_prediction
This Python project uses the PyTorch library to predict the value of a digit represented by an image (28x28x3). These digit images have different colorations.
DeepRL_HighwayEnv
This project trains RL agents on Highway and Racetrack environments using custom DQN and PPO implementations with grid-based observations and tailored reward shaping. Agents learn to stay on road, drive fast, and avoid collisions. A Stable-Baselines3 version with Optuna is also included for benchmarking.
GAN_cat_image_generator
This Python project uses the PyTorch library to generate unique cat images using GAN technology.
Genetic_Algorithm-3dna
The aim of this Python project is to circularize a plasmid by optimizing the values of a table calculating the plasmid's 3D trajectory. To achieve this, 2 algorithms are used and compared: simulated annealing and genetic algorithm.
SAT_music_generation
This project generates harmonized piano music using SAT and Pseudo Boolean solving, enforcing melodic constraints, controlled repetitions and dynamic chord progressions. It also compares the generation with a linear Gurobi solver.
SHAP-Statistics_tabular_dataset_XAI
This project analyzes the interpretability of linear regressions, random forests and GAM using global and local XAI methods. It uses t-values, feature importance and SHAP on tabular datasets to explain model predictions.
Weld_quality_prediction
This project shows that machine learning can effectively predict weld quality, offering an alternative to physical testing for manufacturers. Selected imputation methods such as MICE and prediction models such as XGBoost and Random Forest have produced promising results.
XLMR_BERT_TextClassifier
This project classifies texts from 390 languages using XLM-RoBERTa and BERT, leveraging multilingual transformers for accurate language identification. It explores preprocessing techniques, data augmentation, and model fine-tuning to optimize classification performance.
ErwanDavidCode
A quick presentation
GradCam-CNN_cancer_detection_XAI
This project trains a CNN to detect cancer in histopathology images using the PatchCamelyon dataset. It applies Grad-CAM, saliency maps, and integrated gradients to interpret model predictions. Grad-CAM shows strong alignment with tumor regions, offering more focused visual explanations than other methods.
LLM_embedders_fairness
This is my 6-month AI research internship in Mila (Montreal). The repo will soon be available and open source !
MAS_robot_navigation
This project simulates a multi-agent system of autonomous robots tasked with clearing radioactive waste in a grid environment. Agents operate under zone-specific constraints and follow different exploration and coordination strategies without communication. The goal is to compare strategies for efficiency in waste collection and merging.