Louis Arts

Louis Arts

MSc Data Science & Machine Learning · UCL

About

I recently graduated with an MSc in Data Science and Machine Learning from UCL, where my thesis, written in collaboration with Chattermill, introduced LLM-in-the-loop methods for clustering and evaluating text without ground-truth labels. I also co-authored research on cross-lingual factual recall in large language models, now under review at EMNLP 2026. Before moving into machine learning I studied astrophysics at Imperial College London and built NLP data pipelines at Lemon AI, a London startup. I care about building genuinely useful models and systems.

News

  • Jun 2026 Started a new position as Research Intern at Chattermill, working on LLM-in-the-loop text clustering algorithms.
  • Jun 2026 Submitted research on cross-lingual factual recall in large language models to EMNLP 2026 (under review).
  • Dec 2025 Published a Medium article about my passion project: building a mobile eye-tracking system from scratch, training a three-branch CNN on 1.5M+ frames to reach 1.4 cm mean gaze error.
  • Sep 2025 Began the MSc in Data Science and Machine Learning at University College London.
  • Oct 2023 Graduated with an MSc in Physics (Astrophysics) from Imperial College London, completing a research thesis on galaxy simulations for objects observed by the James Webb Space Telescope.

Projects

  • CLEAR text clustering pipeline

    Meaning Over Geometry: LLM-in-the-Loop Text Clustering and Evaluation

    MSc thesis in collaboration with Chattermill on criterion-driven text clustering. Introduced the codebook judge, a label-free LLM evaluator reaching a median Spearman correlation of +0.87 with benchmark labels across 19 corpora, and CLEAR, an LLM-in-the-loop algorithm that selects and repairs candidate clusterings, beating the best published result on 8 of 19 benchmarks at roughly one cent per 100 documents.

    Python · LLMs · Text Clustering · LLM-as-a-Judge · NLP
    PDF
  • PolyFact cross-lingual factual recall

    Cross-Lingual Factual Recall via Reinforcement Learning

    Benchmarked cross-lingual factual recall across 100K Wikidata facts and 12 languages; led mechanistic interpretability analysis identifying language-specific MLP neurons and showing GRPO training shifts processing to deeper, more general layers. Under review at EMNLP 2026.

    Python · Mechanistic Interpretability · LLMs · NLP · GRPO
    arXiv GitHub
  • Eye-tracking CNN architecture

    End-to-End Mobile Eye-Tracking System

    Three-branch CNN trained on 1.5M+ augmented frames with custom data preprocessing, achieving 1.4 cm MAE on a held-out test set and deployed as a real-time mobile prototype.

    Python · PyTorch · Computer Vision · Mobile
    GitHub
  • Facial emotion recognition

    Real-Time Facial Emotion Recognition

    Transfer learning on DenseNet169 with data augmentation, achieving 81% test accuracy on a class-balanced dataset of 40K+ labelled images. Presented at Le Wagon Demo Day.

    Python · TensorFlow · Transfer Learning · OpenCV
    GitHub