A collection of all my projects, including Data analysis, Data Science, Python development using Django and End-to-End LLM Projects
Data enthusiast, Python programmer, worked on creating apps using NLP libraries and LLMs for personalized user experiences. Last year student in a double degree master's program in Data Science and AI for competitive manufacturing at ECN France and UNI-T Italy
Built a Model Context Protocol server enabling natural language querying over relational manufacturing databases (SQLite/PostgreSQL/MySQL); Claude autonomously discovers schema, generates SQL, and returns structured analytics, packaged as an installable Claude Desktop extension (.mcpb) with support for user-configured production database connections.
Fine-tuned a multilingual transformer (XLM-RoBERTa) to detect disease mentions in both English and French clinical text, trained on two real medical corpora. Built a second pipeline that first identifies whether a document is medical, legal, or financial, then runs the appropriate NER model to extract relevant entities from it. Also tested how well the English-trained model generalizes to French with zero additional French training data.
Hybrid regex + spaCy NER pipeline that detects and redacts PII from English and French documents across medical, legal, and financial domains. Supports three redaction strategies: full masking, typed placeholders, and consistent surrogate replacement using locale-aware fake data generation.
End-to-end document understanding pipeline processing receipts and tables from raw images. Uses Donut for zero-OCR receipt extraction (company, date, address, total) and a two-stage Table Transformer + Tesseract approach for structured table parsing. Evaluated across 50 documents with per-field accuracy metrics.
AI-powered Database Query Assistant (LangGraph + LLMs) that converts natural-language queries into SQL, executes them on PostgreSQL, and generates concise analytical insights.
n8n workflow that automatically fetches trending GitHub repositories related to AI/LLM topics, processes the data, stores it in a Postgres database, and sends a weekly summary email with the top repositories.
Local RAG chatbot enabling document analysis through natural language queries, built with LangChain, ChromaDB, and Ollama for privacy-focused offline operation. Implemented modular architecture with conversational memory, async processing, and source attribution - processing PDFs and articles with 800-token semantic chunking and vector similarity search
A Resume ranking system, comparing Job Descriptions with the text provided in the Resumes by utlizing basic to advanced NLP models
Classifying defaults and non defaults using a dataset which contains customer financial profile parameters, like credit history length and account age etc. The solution includes application of several ML models and hyperpramater tuning to find the best model and obtain best accuracy.
Implemented all ML models along with advancded algorithms like XGBoost, Deep learning algorithms, on industrial historical data to analyze and predict the production delays in orders
A Knowledge Graph that integrates transportation networks and educational facilities across the Trentino Province. It applies iTelos KGE principles and FAIR guidelines to ensure data reusability, interoperability, and semantic consistency. The resulting graph enables users to discover optimized commuting routes by combining real-time transport data with structured educational location information
The Project involves Exploratory Data Analysis (EDA) and insights extraction from a complex dataset to understand the impact of price changes on sales. It includes applying forecasting algorithms and visualizing results using Tableau.
An end-to-end sales analytics project using SQL for data transformation and Power BI for dynamic dashboarding. Showcases KPI engineering, customer insights, and interactive data visualizations.
This project explores performance optimization in large-scale matrix multiplication by implementing and comparing MPI, OpenMP, and hybrid parallelization approaches. It aims to analyze time and memory improvements across distributed, shared, and mixed memory architectures.
A Q-Learning–based reinforcement learning agent trained to navigate a 10x10 grid, reach the treasure, and avoid dynamic obstacles using Stable Baselines3. Achieved a 76% success rate with an average reward of 13.54 and 8.84 steps per episode.