Work

Solo-built platform with nine AI agents that coordinate across the full data science pipeline. Humans in the loop at every decision point. Live with data scientists at Fetch.ai.

Detecting condescending language at a 9.5:1 class imbalance. By hand, without AI support as the department required, a DeBERTa-v3 model with a community-aware prefix and focal loss: F1 0.59 on the dev set against a 0.48 baseline, 10th of 340, 29 of 30 for the report. Then my agent system on GPT-4o came 22nd unsupervised and 1st with me in the loop.

MEng thesis. A controlled backbone swap on 4,128 fetal ultrasound subjects, training no new model. Holding the head fixed, domain pretraining adds 0.053 to 0.086 AUROC, and FetalCLIP with a transformer head reaches 0.911. At a 0.20 screening threshold that model does worse than referring nobody on all five folds, while the DINOv2 linear probe does better on all five. The model that wins the ranking loses the decision.

An operating system in C and a compiler in Scala, both from scratch, in second year. 91.19 across the 15-credit module. The year after, I was paid to teach it: three two-hour labs a week, twenty students.

Fine-tuned open-source TTS to express emotion. Compiled an emotional audio dataset with 7,000+ HuggingFace downloads.

Evaluating LLMs across 70 languages and 21 families. Found that Tamil and Malayalam cost 10 to 12 tokens where English costs 2 to 3, so a per-token price is a tax on the languages it is hardest to serve.