Active Learning on CIFAR-10
ResNet embeddings, TypiClust / typicality-based selection, semi-supervised training and accuracy-vs-budget analysis — the full experimental pipeline in one notebook.
Written reports and research notebooks behind the projects — the technical detail, methods and results, not just the headline.
ResNet embeddings, TypiClust / typicality-based selection, semi-supervised training and accuracy-vs-budget analysis — the full experimental pipeline in one notebook.
Transformer-based classification of clinical text for the mental-health assistant, with EDA and medication/therapy prediction notebooks alongside.
Data visualisation, feature construction and a controlled comparison of models on tabular data, with a dedicated evaluation notebook.
The multi-agent test-generation pipeline (Programmer → Tester) benchmarking self-hosted SLMs against Claude on HumanEval & MBPP, with the run artefacts behind the reported numbers. A walkthrough demo is on the Projects page.