Quadratically Large Regular Graphs Without Internal Partitions
Quadratic-size counterexamples to Ban and Linial's proposed linear threshold for internal partitions.
Read paper →
/// What I Do
Distributed Systems
Building scalable systems for ambitious products. Experienced in creating high-performance infrastructure that scales to billions of requests.
AI & Reinforcement Learning
Creating AI agents that master complex games. Built environments for Terra Mystica, SpeedRunners, and research into multi-agent learning systems.
Full-Stack Development
From React Native mobile apps to Rust-based game engines. Comfortable across the entire stack with a focus on performance and user experience.
/// Papers
Quadratically Large Regular Graphs Without Internal Partitions
An internal partition splits a graph into two nonempty groups so that every vertex has at least as many neighbors on its own side as across the split. These constructions grow quadratically with degree, disproving Ban and Linial's proposed linear threshold.
Read paper →
/// Featured Projects
RDF
Ontology
Prophet
Ontology-first backend scaffolding across Java, Node, and Python
A multi-stack code generation platform that compiles a single domain ontology into SQL, OpenAPI, and framework-specific integration code. Prophet reduces contract drift by generating typed actions, query APIs, persistence adapters, and migration artifacts with deterministic outputs and compatibility checks.
View Project Details →
TypeScript
React Native
Splice
AI-powered bill splitting mobile app
A React Native mobile app for seamless expense tracking with AI-powered receipt scanning. Built with TypeScript, Expo, and Supabase, it features real-time sync, passwordless authentication, and intelligent OCR to automatically parse receipts and split bills among groups.
View Project Details →
Rust
Python
Terra Mystica RL
High-performance board game AI environment
A rules-complete implementation of Terra Mystica in Rust, exposed as a PettingZoo/PyTorch RL environment via PyO3. The core engine leverages Rust's performance guarantees for maximum simulation speed, while Python bindings provide seamless integration with modern RL frameworks. Includes Pygame visualization for debugging agent behavior.
View Project Details →
Python
PyTorch
HLRL
Modular reinforcement learning library
A from-scratch RL library implementing state-of-the-art algorithms including SAC, DQN, IQN, RND, Ape-X, and R2D2. Features a flexible wrapper-based architecture that allows algorithms to be composed and mixed, with backend-agnostic core abstractions for portability. Optimized for single-machine high-performance training without sacrificing code clarity.
View Project Details →
Python
C++
SpeedRunners RL
Deep RL agents playing SpeedRunners via direct game injection
A modular reinforcement learning system that trains agents to play SpeedRunners using direct game process injection. A C++ DLL hooks into the game to extract state and inject actions, exposing a Gymnasium-compatible environment via named pipes. Agents are trained using Rainbow IQN and RND in PyTorch, with a clean separation between game interfacing (sr-lib), environment wrapping (sr-gym), and agent training (sr-ai).
View Project Details →