Projects
A mix of research and personal work, mostly around systems programming, machine learning, and high-performance computing. Write-ups for some of these live on the blog.
C++, SFML
A modular 2D physics engine from scratch in C++ (rigid body dynamics, velocity-based movement, AABB collision detection), extended into an orbital mechanics simulation for gravitational N-body interactions via numerical integration, with a real-time SFML visualization layer.
Active development · Python, gRPC, FastAPI, React
A distributed inference system for state space models (Mamba, S4) and small LLMs across a 6-node Raspberry Pi 5 cluster. Handles node discovery, model sharding, and pipeline-parallel execution transparently: each Pi runs a daemon that broadcasts presence, while an orchestrator maintains a live node registry and dispatches work over gRPC. Accessible through a Textual TUI and a React WebUI.
C++, PyTorch
Monte Carlo Tree Search combined with a CNN for board evaluation and move prediction. Most of the effort went into trimming the model, from 85MB to 42.66MB (50% smaller), while making it 10× faster per move (1000ms → 100ms) and keeping a >90% win rate against the larger baseline.
C, xv6 kernel
Could a modern LLM inference engine run on a minimal educational OS? A POSIX-compliant shared-memory subsystem for the xv6 kernel caches model weights (~100MB) without redundant transfers; custom threading primitives, built from scratch, parallelize the inference engine itself. End result: 16.2 tokens/sec, up from ~7.
Python, XGBoost, LightGBM, CatBoost
A stacking ensemble of XGBoost, LightGBM, and CatBoost with a logistic regression meta-learner, combined with preprocessing for extensive missing data and non-predictive features. Finished 1st on imbalanced tabular data (Private Leaderboard AUROC: 0.64671).
C++, SFML
A modular 2D physics engine from scratch in C++ (rigid body dynamics, velocity-based movement, AABB collision detection), extended into an orbital mechanics simulation for gravitational N-body interactions via numerical integration, with a real-time SFML visualization layer.
Active development · Python, gRPC, FastAPI, React
A distributed inference system for state space models (Mamba, S4) and small LLMs across a 6-node Raspberry Pi 5 cluster. Handles node discovery, model sharding, and pipeline-parallel execution transparently: each Pi runs a daemon that broadcasts presence, while an orchestrator maintains a live node registry and dispatches work over gRPC. Accessible through a Textual TUI and a React WebUI.
GitHub
Pulled live from the GitHub API.
Loading repositories from GitHub…