Tom Genlis

End-to-end project that learns to predict short-term 3D vehicle translations from sequential unlabeled LiDAR point clouds from SemanticKITTI. Combines a PointNet-style encoder, an LSTM temporal model, balanced sampling, and a composite distance + direction loss.

EF Mistral AI Game Jam. Mysterious train with locked wagons. Players gather clues from passengers and provides them to an LLM detective that guesses the wagons' passcodes. The game integrates LLMs, TTS and Unity.

Developed an AI app running fully on-device, specialized in answering museum-related questions and recognizing artworks. Fully on device LLM (Llama 3 1B) running on iPhone, specialized in l'Orangerie gallery in Paris (using RAG and memory).

Efficient and clean reimplementation in PyTorch of ground breaking Deep RL paper Rainbow DQN (Hessel et al.) for Atari Gym environments. Includes a paper-like report.

Photo recognition mobile app for identifying different dog breeds. Developed a ViT transformer model with 92.3% accuracy.

Created dynamic Slurm bash utilities for job management, tailored to Siemens' high-performance computing cluster.