Selected work
A selection of projects I've built, machine learning models, computer vision systems, data pipelines and web applications. All of them are open source on GitHub.
An AI answer engine built as a single Next.js app. Gemini streams answers grounded in live Google Search results, shown with numbered citations, while Supabase handles sign-in, chat history under row-level security and per-user rate limits.
Compares a classical pipeline (CLAHE preprocessing, GLCM, LBP and HOG features, PCA, SVM and ensemble classifiers) with a ResNet18 fine-tuned by transfer learning, and explains the network's predictions with Grad-CAM heatmaps. An educational project, not a clinical tool.
A content-based music recommender in Python and Streamlit. Enter a song and get similar tracks by cosine similarity over weighted audio features, blended with popularity through a slider, with an optional genre filter. Similarities are computed on demand instead of stored as a full matrix.
Detects faces with OpenCV's DNN module and classifies each into an age range with a pretrained Caffe model, from an image file or a live webcam, drawing the result on the frame. It predicts ranges, not exact ages.
Super resolution project that upscales image quality by 2–5×.
Tracks how Hacker News talks about the major AI companies. A collector keeps only AI-related front-page stories and tags each with the company it is about; a Streamlit dashboard then runs sentiment analysis per company.
A semester project for an Internet Technologies course: a session-protected, admin-only student-records portal built with Node.js, Express, EJS and SQLite. It also exposes the data as XML transformed into HTML with XSLT, and adds CSRF protection, rate limiting and multi-language support.
A clean, minimal Gemini API implementation in Python.