Enguerrand Le Niliot

Multimodal Product Classification (Rakuten France)

ENS Challenge Data, 2026

1st of 135 on the public leaderboard, weighted F1 0.9346 (October 2026)

Predict the product type (27 classes) of about 99,000 Rakuten France listings from their French title, description and image. I built a stacking pipeline on one shared 5-fold split: TF-IDF word and character n-gram models, then a meta-model (logistic regression and gradient boosting) trained on their out-of-fold scores. The largest gain came from an exploratory analysis of how the catalogue was assembled, turned into features and checked with a strict validation in which no validation label is used. The pipeline also includes fine-tuning scripts for French text transformers and image backbones (ConvNeXt, SigLIP 2).

Next-Job Recommendation from Browsing Sessions (HrFlow.ai)

ENS Challenge Data, 2026

1st of 69 on the public leaderboard, score 0.7354 (October 2026)

From the sequence of job ads a candidate opened, predict the next 10 ads they will open (MRR@10) and whether they will view or apply. I built a two-stage recommender: candidate generation (co-visitation, Session-kNN, text kNN, SVD embeddings) followed by a LightGBM learning-to-rank ensemble on about 80 features. An exploratory analysis of how the sessions were built led to a dedicated classifier (AUC 0.990) whose probability is fused with the ranker scores. Validation is leak-free: GroupKFold by profile, with every target-dependent statistic recomputed out of fold.

Trust or Bet Against? Asset Allocation Returns (QRT)

ENS Challenge Data, 2026

39th of 1,702 on the public leaderboard, accuracy 0.688 (official benchmark 0.508)

Predict the sign of the next-day return of systematic asset allocations from 20 days of history. Market signals alone are weak (0.522 at best, mostly short-term momentum), so most of the work went into reconstructing the hidden structure of the dataset from cross-sectional return correlations and group-level PCA factors, with every step measured on simulations before use. Models are LightGBM, validated with purged time-block cross-validation.

Language-Conditioned Visual Navigation (RoboWorld 2026, WorldNav)

Challenge of the RoboPAD workshop at NeurIPS 2026, in progress (final phase in November 2026)

2nd on the phase 1 leaderboard, score 53.8 (organizers' baseline 20.4)

From a single egocentric RGB image and a natural-language instruction, predict the full sequence of robot actions. My model fuses tokens from pretrained vision (SigLIP 2, DINOv2) and text (BGE) encoders in a transformer that outputs 12 trajectory hypotheses anchored on k-means prototypes, which avoids the mode collapse of plain winner-takes-all training. Hypotheses are decoded per data source by minimum Bayes risk on the official metric, which I re-implemented locally to tune the decoding.

LoL Draft Advisor: Win Probability for League of Legends Drafts

Personal project, 2026

54.4% accuracy on 23,000 unseen recent matches, well calibrated

A draft assistant that ranks the champions available for your role by the win probability they give your team, at any point of the pick phase. I wrote a rate-limited Riot API crawler that collected over 150,000 ranked matches into SQLite, then an additive logistic model with one regularisation strength per group of effects: champion and role strength, synergies, team damage balance, and lane matchups learned from a continuous auxiliary target (gold share) with cross-fitting. Champion-specific effects are shrunk with empirical Bayes. Draft alone explains little of a solo-queue result, and the accuracy is close to the known ceiling for this task; negative results (factorization machines, a BERT-style pick model) are documented. Served as a Streamlit app.

White Blood Cell Classification

Télécom Paris, IMA205 course challenge, 2025

5th of 77 submissions, macro F1 0.7839

Classify 13 types of white blood cells from microscopy images with strong class imbalance and inconsistent staining. A classical pipeline (segmentation, hand-crafted features, SVM) served as a baseline, then I fine-tuned a ConvNeXt with focal loss, weighted sampling, CLAHE preprocessing, 4-fold stratified ensembling and test-time augmentation.

Unsupervised Clustering of Abstract Paintings

Télécom Paris, IM06 research project, 2025, team of 4

Group abstract paintings by style without labels, a setting where standard CNN features fail because there are no objects to recognise. I was responsible for the CNN style features (Gram matrices over 4 VGG16 layers, random crops and PCA), their fusion with VAE and geometric descriptors, and a silhouette-based filtering step that removes ambiguous paintings from K-Means clusters before re-clustering them. The clusters are explored in an interactive WebGL map.

Bubble Battle (Global Game Jam 2025)

48-hour game jam, team project

A local two-player fighting game where each player is a bubble bouncing around a closed arena and uses momentum-based abilities to knock out the opponent. I programmed the player movement and physics on Godot's CharacterBody2D, designed the four arenas and balanced the four abilities, mainly through their cooldowns.