Research

Observable language models: every hidden state is a readable distribution over words (hidden dim = vocab size). A uv workspace of architectures plus a glassbox CLI for tracing, causal probing, and faithfulness coverage, chasing observability at no performance cost.

BCS Thesis: Reinforced Linear Genetic Programming

A deep learning system combining transformers and CNNs to classify diseases from patient-described symptoms, achieving 90% recall through semantic embeddings and CNN feature extraction.

A CNN research project exploring optimal image classification architectures for small datasets, demonstrating that data quantity remains the fundamental bottleneck for deep learning performance.