Projects

Research experiments, contest work and apps. Source links where public.

Tacit Collusion in LLM Bidders

48-session baseline · 2 models · 4 arms

Conceived the study and built its auction harness: does letting LLM firms communicate raise prices? Four pre-registered arms: no channel, private notes, a shared workspace, an open message channel.
Ran a pre-registered no-channel baseline of 48 sessions (2 models, 12 seeds, 50 rounds), each matched with a one-shot control. With history, DeepSeek V4.1 Flash bid about 1 to 2.3 below the Bayes-Nash bid by round 50 and GPT-6 Luna about 6 above it; neither control drifted.
Merged a collaborator's shared-workspace and message-channel code, then built the private-notes control as a one-switch variant of it, with a log check that no prompt shows a firm the other's notes.
Pre-registered two hypotheses in configs/analysis.yaml before any session ran (paired exact sign-flip tests, Holm-corrected) and ran 24 GPT-6 Luna notes sessions. Analysis in progress.

Aug 2025

Bronze · 99.62 top task score

Placed 2nd on a segmentation task with a PyTorch model that handled a 50× class imbalance, raising accuracy on the critical human/wall classes by 37% through data augmentation and tuned CrossEntropyLoss weights.
Solved 6 machine learning tasks (NLP, CV, deep learning) solo in a 2-day contest under 12-hour time and memory limits.
Delivered a pick-and-place pipeline for the Galbot G1 robot with a team in a MuJoCo simulation, using A* path planning and inverse kinematics to navigate obstacles and organise objects on shelves.

2nd + Innovation Award · JA Code for Impact

Built a full-stack React Native app with a team to help users discover environmental initiatives, with an AI assistant and a carbon footprint calculator.
Cut average post load time by 46% for 100+ concurrent users (Supabase queries, client-side caching) and Android build times by 65% (Expo and Gradle).

96% validation accuracy · 11,000+ images

Reached 96% validation accuracy on 11,000+ colonoscopy images by fine-tuning ResNet18 with augmentation, weighted sampling and staged unfreezing.
Built a Flask app for multi-image upload and real-time inference that aggregates predictions into a risk rating, with the model traced to TorchScript for CPU deployment.