Data Science & AI undergraduate at Nanyang Technological University (NTU), on the Turing AI Scholars' Programme (graduating 2028). Aspiring quant — interested in AI, game theory, and probabilistic modelling.
PoE-World: Compositional World Modeling with Products of Programmatic Experts · builds on topwasu/poe-world
- Reproduced and extended PoE-World, a symbolic world model representing Atari dynamics as a Product-of-Experts over 400+ LLM-synthesised programs, with weights fit by L-BFGS optimisation.
- Swapped the synthesis backbone from GPT-4o → DeepSeek-Chat, cutting per-run API cost by ~95%, and added defensive guards against hallucinated object types that were crashing the L-BFGS optimiser.
- Evaluated across three games: reproduced Pong within run-to-run variance, but found planning collapsed on Breakout and Montezuma's Revenge — tracing both to hallucinated object dynamics, quantified across all synthesised experts.
Code lives in a private NTU lab repository; available on request.
- pushfold-solver — GTO push/fold solver for short-stack NLHE: exact & Monte-Carlo equity, Nash equilibria (2/4/9-handed), a stochastic-approximation solver, and 242 tests.