Tristan Donzé
Final-year Research Master’s student in ML
Institut Polytechnique de Paris
I'm a final-year Research Master’s student in Machine Learning at Institut Polytechnique de Paris (École Polytechnique). You can find more about my work and projects here. I also have a blog where I share my thoughts on AI, research papers, and project updates.
Education
Institut Polytechnique de Paris
M.S. in Artificial Intelligence
Research Master’s in AI (English-taught) | Courses delivered at École Polytechnique & Télécom Paris
University Paris 8
B.S. in Computer Science
Experience
Freelance AI Engineer — Histia
Developed an autonomous multi-agent pipeline for analyzing company websites and generating investment-grade reports. Processed textual content (markdown) and visual context (screenshots) for complete page understanding. Merged page-level summaries into unified insights describing products, strategy and positioning. Implemented a coordination mechanism to direct navigation based on missing report information. Designed to automate analyst workflows for VC funds, incubators and M&A advisory teams.
Applied Research Intern — Histia
Designed an end-to-end system for company logo detection and identification in natural images. Prepared and standardized a ~3M instance dataset with a real-world annotated test set. Reviewed state-of-the-art methods and fine-tuned CLIP with contrastive learning and LoRA. Improved performance through iterative error analysis and targeted data augmentations. Achieved 94.6 percent Top-1 accuracy, surpassing the baseline by more than ten points.
Portfolio
Deep Reinforcement Learning Blackbox Challenge
Built and trained RL agents to solve a fully unknown environment with highly noisy observations and no access to dynamics. Implemented PPO and A2C with GAE, entropy scheduling, KL early stopping, and cosine annealing. Designed a compact shared encoder to maximize representational efficiency under strict model size limits. Achieved top performance in the challenge.
Political Speech Imitation and Fallacy Detection
Fine-tuned multiple LLMs using QLoRA to imitate political speech styles and detect logical fallacies across large argument corpora. Trained models on 16k rhetorical instruction pairs and evaluated fallacy prediction across 9 reasoning error types. Analyzed stylistic shifts, fallacy frequency, and per-class model performance.
Football Event Detection from Tweets
Detected key events in World Cup matches from multilingual Twitter streams by classifying short time windows during the match. Experimented with BERT fine-tuning, embedding-based classifiers, a MLP using top-n word importance matrices and a Temporal Convolutional Network. The best performance was obtained with the TCN and the MLP.
News
September 2025
Second place at {Tech: Europe} Paris AI Hackathon!
We built MagnOSS, an AI-powered chess learning platform, and secured second place among 30+ teams at the {Tech: Europe} Paris AI Hackathon!