Software, machine learning and deep learning projects

Four chest X-rays from the test set and, below them, their Grad-CAM++ maps for emphysema, with the model's attention focused on the lung fields.

01 / Medical Imaging

Multi-label chest X-ray classification with DenseNet-121

Detecting 14 pathologies at once in 112,000 chest X-rays, with an asymmetric loss, per-pathology calibrated thresholds and Grad-CAM++ to audit the model.

Four greyscale skin lesions. Next to each one, the same image with the ground-truth mask outline in amber and the model's prediction outline in blue.

02 / Medical Imaging

Melanoma segmentation with U-Nets built from scratch in PyTorch

Two U-Nets coded from scratch to outline skin lesions: one trained from scratch and one with a frozen pretrained ResNet-50 encoder.

On the left, twenty-four CIFAR-10 images (ships, dogs, horses, cars) enlarged as pixels; on the right, four handwritten MNIST digits.

03 / Computer Vision

Ablation study on MNIST and CIFAR-10: 99.47% and 93.87% accuracy

Two ablation studies measuring how much each piece contributes to beating 99.4% on MNIST without convolutions and 92% on CIFAR-10.

Diagram of the app in three cards. An audio waveform goes into Whisper, which produces transcript lines, and then into a 4-bit LLM that writes a short summary. The last card lets you choose between Qwen 14B and Phi-3 mini.

04 / Generative AI

Local video transcription and summarisation with Whisper and a quantised LLM

A web app that transcribes and summarises videos with Whisper and a quantised 14B LLM, all locally and with no external APIs.