Research on privacy-preserving machine learning with homomorphic encryption at Zama, as part of a final-year project in collaboration with this renowned company specializing in open-source cryptographic tools and homomorphic encryption through Concrete ML. This project involved exploring practical applications of privacy-preserving machine learning (PPML) while working in a 6-person team dedicated to implementing and evaluating models that protect data confidentiality without compromising on performance.Key implementations included developing and testing machine learning models with Concrete ML for real-world applications such as predicting CV relevance for JavaScript developer roles and detecting banking fraud. These use cases allowed us to analyze the feasibility of FHE-based models by comparing their performance and accuracy to traditional models in scikit-learn, all while ensuring data privacy through homomorphic encryption.We also deployed a secure fraud detection model on Hugging Face with API accessibility, demonstrating how Zama can offer a concrete privacy-preserving machine learning solution to users.Technologies and skills applied in this project included a strong command of machine learning with tools such as PyTorch, NumPy, and Concrete ML; extensive experience in privacy technology through homomorphic encryption and PPML methods; programming proficiency in Python; and solid research capabilities in benchmarking, performance optimization, and designing machine learning models focused on privacy

Achieved a ~70% improvement in RSSI prediction error between LoRaWAN gateways and devices, reducing it from approximately 15 to 4. -Designed and implemented a terrain classification tool using Google Maps API, resulting in enhanced precision in network simulation. This tool played a pivotal role in refining our approach to network layout and coverage.-Handled geospatial data processing and integrated Mapbox tiles for high-resolution satellite image analysis in urban projects, providing crucial data insights that significantly contributed to the enhancement of the RSSI prediction algorithm.-Conducted comprehensive spatial data analysis at a city-wide scale, encompassing hundreds of thousands of buildings, to assess building density. This extensive analysis was pivotal in devising improved network coverage strategies, directly contributing to the significant enhancement in RSSI prediction accuracy.