Projects
Full project details
The same information as the portfolio project cards, available as plain HTML.
2026
Low-Dose CT Reconstruction & Tumor Segmentation
Work in progress
Overview
An ongoing medical imaging project investigating the impact of simulated low-dose CT reconstruction on liver and tumor segmentation.
Keywords
Medical Imaging · CT Reconstruction · Deep Learning · Image Segmentation · U-Net · Filtered Back Projection · Total Variation · Poisson Noise · PSNR · SSIM · Dice · HD95
Approach
- Developed a medical imaging pipeline to investigate the impact of simulated low-dose CT reconstruction on liver and tumor segmentation.
- Implemented Filtered Back Projection (FBP) and Total Variation (TV) regularized reconstruction with Poisson-based photon noise simulation.
- Trained a 2D U-Net for liver and tumor segmentation and evaluated reconstruction quality (PSNR, SSIM) and segmentation performance (Dice, HD95).
2026
Local RAG Chatbot for LIP6
Overview
Developed a fully local, end-to-end RAG chatbot for the LIP6 institutional website, covering data ingestion, retrieval, generation, evaluation, and serving. The system processes heterogeneous HTML and PDF content, structures and indexes it locally, retrieves relevant sources through hybrid dense and sparse search, and generates cited answers using locally hosted open-weight LLMs.
Keywords
RAG · LLMs · Information Retrieval · Hybrid Search · Embeddings · Vector Databases · NLP
Collaborators
Approach & Results
- Conducted a literature review of state-of-the-art RAG architectures and built a fully local pipeline for ingesting, structuring, and retrieving information from LIP6 HTML and PDF content.
- Implemented hybrid retrieval by combining dense search with Qdrant, sparse search with FTS5, and reciprocal rank fusion.
- Compared chunking strategies, embedding models, retrieval configurations, and locally hosted LLMs, and deployed ingestion and inference on an NVIDIA DGX Spark.
- Evaluated retrieval quality, latency, and citation accuracy using external benchmarks and a curated Golden Questions dataset.
- On a 79-question robustness benchmark, achieved 86.1% Recall@10 and 77.2% Hit@5 across challenging real-world queries.
- 72.6% of generated answers cited the expected source, while 78.5% included at least one citation.
2026
BirdCLEF+ 2026 - Bioacoustic Species Classification
Overview
Developed a classical machine-learning pipeline to identify animal species from audio recordings and noisy, multi-species soundscapes. The task covered more than 200 species, severe class imbalance, and both single-label and multi-label classification. Under the constraint of not using deep-learning models, the project focused on handcrafted audio features and traditional machine-learning methods.
Keywords
Audio Classification · Machine Learning · Multi-label Classification · Signal Processing · One-vs-Rest · Feature Engineering · Data Augmentation · Class Imbalance
Collaborators
Approach
- Engineered a 490-dimensional audio representation combining MFCC-based and spectral features with statistical summaries.
- Addressed rare species and heterogeneous recordings through normalization, duration handling, and targeted audio augmentation.
- Compared Logistic Regression, SVC, and XGBoost using One-vs-Rest classification, cross-validation, and hyperparameter tuning.
- Implemented whole-recording feature aggregation and examined alternative strategies, including overlapping five-second chunk classification, for better capturing temporal information under additional time and computational resources.
2026
Sentiment and Speech Classification
Overview
Built and evaluated NLP pipelines for two binary text-classification tasks: speaker identification from French presidential speeches (Chirac vs. Mitterrand) and sentiment classification of English movie reviews. The project compared text preprocessing, representations, classical machine learning, recurrent networks, and transformer-based models, exploring how textual and sequential features support supervised classification.
Keywords
NLP · Text Classification · Sentiment Analysis · Topic Modeling · Text Embeddings · Classical ML · RNNs · Transformers · Fine-tuning
Collaborators
Approach & Results
- Analyzed vocabulary, class distributions, and topics using EDA, LDA, and LSA.
- Compared BoW, TF-IDF, Word2Vec, FastText, Doc2Vec, and E5 embeddings with classical classifiers.
- Evaluated bidirectional GRU/LSTM architectures and fine-tuned CamemBERT, BERT, and RoBERTa models.
- Explored sequential post-processing for speaker predictions; transformers raised F1 from 0.62 to 0.81 for speaker identification and from 0.81 to 0.90 for sentiment classification.
2025
Explainable Board-Game Recommendation System
Overview
Developed an explainable recommendation system for board games using TricTrac user ratings, reviews, and game descriptions. The project compared collaborative-filtering and content-based approaches, then used NLP to make recommendations more interpretable through cluster explanations and personalized, review-style recommendations.
Keywords
Recommender Systems · Explainable AI · Collaborative Filtering · Matrix Factorization · Clustering · NLP · Text Embeddings · LLMs
Collaborators
Approach & Results
- Compared k-NN collaborative-filtering variants on a 98.1% sparse user-item matrix, using rating centering and distance weighting and evaluating performance with RMSE and MAE.
- Applied NNMF for dimensionality reduction and interpretable latent-factor extraction, selecting 20 factors through cross-validation.
- Grouped 2,614 games into 30 clusters using K-means and visualized their structure with t-SNE.
- Explained clusters using game-description embeddings, frequent bigrams, and LLM-generated summaries.
- Generated recommendation explanations from approximately 96,000 reviews using TF-IDF, n-grams, and sentence embeddings.
2025
Decoding the Bottle - Wine Rating Prediction
Overview
Collected wine data from Vivino and SimpleWine to study how price, origin, grapes, and taste profiles relate to public ratings. The project explored dataset biases, imputed missing values, and compared regression and classification models to predict ratings and identify wines rated above average within their price range.
Keywords
Data Science · Regression · Classification · CatBoost · SHAP · Data Collection
Collaborators
Approach & Results
- Collected, harmonized, and explored wine data from two sources, investigating biases and imputing missing values.
- Compared Random Forest, XGBoost, and CatBoost models; CatBoost performed best among the tested configurations.
- Achieved a held-out MAE of 0.163 stars and R² of 0.553 for public-rating prediction.
- Achieved F1 of 0.684 and 66.7% accuracy for predicting above-average ratings within a price range.
- Used SHAP to interpret model predictions, with price emerging as the strongest driver of predicted ratings.