High-Elo chess bot that pairs policy learning from human games — an HSTU sequential transformer pretrained on Lichess — with value-based search (a WDL value head plus depth-2 policy-pruned minimax). Full build log in Building a Chess Bot with HSTU.
Python library for generating fixed-dimensional embeddings from multi-vector representations, based on Google Research’s MUVERA. C++ core exposed to Python via nanobind and published on PyPI.
Exploring Reinforcement Learning algorithms on customized Lunar Lander environment with dynamic realistic wind vectors by extending the gym environment.
ML inference framework for pytorch models in Asynchronous C++ which supports dynamic batching, Arrayfire, quantized model, and various optimizations written using Drogon
