Eight projects in reinforcement learning, generative modelling, self-supervised learning and systems engineering. Each page covers the question, the method, and what the results actually showed.
Four generative model families implemented from scratch and compared under optimal-transport metrics, isolating the trade-off between sample quality and the number of network evaluations needed to draw a sample.
A cooperative multi-agent framework that uses a frozen LLM as a training-time macro-planner, then distils the guidance into the policy so the LLM can be removed entirely at deployment.
A mixture-of-experts agent for the LevDoom Seek and Slay benchmark — five level-specialised DQN experts routed by a learned scene classifier, compressed to fit a deployment size budget.
Binary segmentation of grayscale medical images, built to isolate one question: does the improvement come from the architecture, or from the initialisation?
Predicting a Go player's rank from 1D to 9D out of game records at 80.65% accuracy — treating rank as an ordered quantity rather than nine unrelated classes, with predictions pooled across views, seeds and games.
A secure micropayment platform in C++ and OpenSSL, where account operations go through a thread-pool server over TLS and payments themselves travel directly between peers.
A full-stack platform for organising physical meetups, with real-time private and group chat, a friend system and an admin console, on a normalised PostgreSQL schema deployed serverless.
A travel assistant on LINE that collects preferences through multi-step dialogue and generates personalised itineraries with Gemini, backed by an OAuth-protected analytics dashboard.