Projects
They trusted me
LLM Deployment with Mixtral
Deployed Mixtral large language model on AWS infrastructure to ensure confidentiality and internal security compliance.
Brain Image Segmentation
Implemented and compared YOLO, StarDist, Detectron, and CellPose models for microscopic neuron counting and segmentation at Neuro-Sys.
Helicopters anomaly detection
Detection of anomalies such as fretting and corrosion on helicopters’ main gear boxes using deep learning for image segmentation.
Helicopters Dashboard for Fleet Tracking
Built different interactive fleet tracking dashboards using Skywise and Spotfire, integrating live SQL queries on Big Data to monitor helicopter operations in real time.
HE-Stain Domain Transfer
Used CycleGANs for domain adaptation between different HE stain styles on Whole Slide Images.
Fuel Flow Prediction
Prediction of helicopters’ fuel flow consumption during takeoff using tabular data and ML models.
PathLDM - Histopathological Image Generation
Used PathLDM to generate histopathological images from prompts for cancer diagnosis pipeline.
HistoGPT - Report Generation from WSI
Fine-tuned HistoGPT to generate medical reports from Whole Slide Images using image-to-text generative models.
Python flask automated analysis app
Developed a Flask-based web app to conduct automated analaysis on different In Vivo/In Vitro files
AI User Study Platform
Developed a Flask-based web app to conduct performance evaluations of AI models with pathologists.
Transformer POS Tagging
Compared pre-trained transformers with home-made models for POS tagging on benchmark datasets.
Multimodal Turn-Taking Prediction
Built a deep learning model that predicts turn-taking from multimodal signals including video, audio, and text.
Machine Learning for Tabular Data
Developed end-to-end ML pipelines for various tabular prediction tasks, including house price estimation, pet adoption speed classification, fraud detection, and Olympic gold medal forecasting.
Research : Molecular Structure Generation
Collaborated with Aix-Marseille University's chemists to generate valid molecular structures from atom lists using constraint programming in Java with the Choco solver. Ensured compliance with chemical constraints through automated structure generation.
Research: Learning Compact Transparent Models using Neuro-Symbolic Methods
Worked with IBM on improving model interpretability using the R2N approach for automatic feature engineering. Focused on building compact and transparent models through neuro-symbolic methods.
Python Web Application for Live Run Tracking
Developed a web application for the “24h de Peynier” race to enable real-time lap counting by staff for each participant, improving accuracy and efficiency of race tracking.