Lab
Things built to understand something — papers reproduced, systems trained, software written from scratch.
PreviewEntryResultYr
- systemLimit order book matching engineSingle-threaded matching engine in Rust for equity simulation, with integer price representation and a synthetic order-flow generator to drive it.System
market-microstructure, systems, simulationRust26 - analysisPPO against GRPO and DPO on verifiable rewardsPPO with a learned critic, swept across four critic capacities, to find where approximation error tips the advantage toward GRPO's group-mean baseline.Analysis — Schulman et al., 2017
reinforcement-learning, language-models, alignmentPyTorchTransformersSLURM4critic capacities swept26 - reproductionChess engine with C++ bindings and MCTSA C++17 chess engine exposed to Python through pybind11, driving a Monte Carlo tree search and a PyTorch policy-value network.Reproduction — Silver et al., 2017
game-ai, tree-search, systemsC++pybind11PyTorchPygame25 - reproductionMinecraft behavioural cloningResNet-style policy trained on self-recorded MineRL play, predicting 11 binary actions from a stack of 20 frames.Reproduction — Baker et al., 2022
imitation-learning, computer-visionPyTorchMineRLOpenCV25