*

A few projects I've built — some for university, others just for fun or to learn something new.

Signal

Creator & Developer (2026)

A privacy-first iPad app that introduces federated learning and AI privacy in under three minutes. Selected by Apple as one of 50 Distinguished Winners of the Swift Student Challenge 2026 and invited to attend WWDC 2026.

achievements

  • Recognised as a Swift Student Challenge 2026 Distinguished Winner, one of 50 worldwide.
  • Built a sandbox where users draw a digit, test a local CoreML classifier, then federate models across simulated hospitals with a live Swift Charts accuracy plot.
  • Simulated privacy attacks and defenses: intercepted updates, data poisoning, Byzantine filtering and differential privacy.
  • Added an on-device AI assistant using Apple's Foundation Models with streamed responses and graceful fallbacks.
  • Shipped with no backend, no network requests and zero third-party dependencies.

technologies

swiftswiftuicoremlfoundation modelsswift charts

Memory Share

Developer

A full-stack social app for sharing geotagged photo posts, explored through an interactive map.

achievements

  • Built geotagged photo posts with an interactive map and multi-criteria filtering.
  • Enforced post visibility rules with Firebase Security Rules.

technologies

react nativetypescriptfirebasefirestore

Movie Genre Classification with Transformers

Developer

Fine-tuned DistilBERT to classify movies into genres from their plot descriptions, built on a large cleaned IMDb dataset.

achievements

  • Reached 73.1% accuracy across 9 genres, a 10% improvement over TF-IDF baselines.
  • Cleaned 11.1M raw IMDb entries down to 500k training records.

technologies

pythondistilberthugging face

TSB-AI-Vision

Creator & Developer

Object detection system using YOLOv5 and PyTorch to detect maritime markers for Técnico Solar Boat. The project was selected as the best in the recruitment process, and I joined the team as an AI Developer.

achievements

  • Trained a YOLOv5 object detection model to identify four different maritime markers.
  • Captured, curated and labeled real-world image data using CVAT.
  • Developed a Colab training workflow that distributed sessions across multiple accounts to avoid usage limits.
  • Set up a shared Google Drive workflow to persist datasets and checkpoints between Colab sessions.

technologies

pytorchyolov5google colabcvatpythoncudaopencv