Harith Proietti Email & Phone Number

Paris, France

- Developed an LLM-based Named Entity Recognition module as part of a medical assistant to offer patients the most appropriate clinicians based on their queries written in natural language.- Worked on various aspects of state-of-the-art GenAI LLMs (Mistral, LLama...) such as prompt engineering (and versioning), model quantization, inference speedup through frameworks like vLLM.- Created an evaluation pipeline based on 500 synthetic medical queries annotated by MDs.- Industrialized the model as an API service (Docker, GCP, Flask).- Developed an advanced Chatbot interface for clients to use the model based on React.js

- Benchmarked open-source English paraphrasing datasets to train a rephrasing model for the Reverso website.- Developed a training pipeline for an English encoder-decoder LLM (Flan-T5), using PyTorch and Hugging Face. Tested different fine-tuning settings based on the LoRA algorithm.- Experimented with different inference strategies (nucleus sampling, top-k, beam search) for quality generation.- Handled the full pipeline for the new French Rephraser: creation of a dataset of French paraphrases with the GPT-3.5 API, research and implementation of different models (encoder-decoder vs decoder, multilingual vs monolingual), tuning of generation parameters, model evaluation, final integration in the API.

Ville De Paris, Île-De-France, France

- Implemented a model aimed to help credit managers detect if a client is likely to face a series of non-payments, leading to a "collective proceeding" event during the incoming year.- Developed a time series approach using sliding windows, lagged features, and a metamodel based on 12 monthly models. Used PostgreSQL for relational databases, Pandas for feature engineering, XGBoost for the model.

Ville De Paris, Île-De-France, France

- Study of the state of the art of the Transformer model applied to Computer Vision (ViT) - Realization of a 45 min presentation (videoconference) on Vision Transformers and self-supervised learning during the monthly meeting of the international Data Science community of Suez - Implementation in pytorch-lighting of the fine-tuning of a pre-trained ViT model, available on the Hugging Face platform - Benchmarking of the ViT model for classification task on standard databases (ImageNet, Cifar10. .) as well as on a Suez internal database- Use of cloud computing with the Azure Machine Learning platform (Azure ML)

Paris, France

- Integration of Latex content on the mobile application and the website of Groupe Réussite (methods, exercises, annals in PC/ECS2) - Writing of articles (methodology, advice) related to the theme of scientific preparation on the blog of Groupe Réussite - Programming of an eligibility simulator for the competitive entrance exams of the Grandes Écoles in React.js

Master 2 - Mathématiques, Vision, Apprentissage (Mva)

Convex Optimization, Probabilistic Graphical Models, Optimal Transport, Object Recognition and Computer Vision, Generative Modeling.

Msc. Data Science, Master 2 Data Science - Artificial Intelligence

Master 2 : Data Science option at Télécom Paris, including courses shared with the M2 Data Science of IP Paris (X, ENSTA, Télécom, ENSAE).

Physique-Chimie-Mathématiques

3 years in a preparatory class for national competitive exams, allowing entrance to French "Grandes Ecoles" (French engineering schools).

Bac S