Noli Jegeni is a sophomore at Northeastern. Currently he's researching world models and making websites. Contact him here.

A chess engine built from scratch in PyTorch: a 10-block residual network over a 102-plane board encoding with policy, value, and a multi-step "lookahead" auxiliary head that predicts the moves played several plies ahead. Trained supervised on a large corpus of master games to ~49% top-1 move accuracy on held-out games, then paired with batched PUCT Monte Carlo Tree Search and exposed over UCI so it can play in any chess GUI or on Lichess.

A Connect Four engine that teaches itself to play from zero human games — an AlphaZero-style value+policy network coupled to Monte Carlo Tree Search, trained entirely by self-play.

Built a custom Gymnasium environment for PicoPark and trained a Stable-Baselines3 DQN agent to a 100% solve rate across multiple procedurally composed levels.

Real-time multiplayer chess variant with a custom "gravity" mechanic that pulls pieces toward the bottom of the board, with live state sync and a lobby system built on Supabase Realtime.

Reimplemented the Hamiltonian Neural Network for a simple Pendulum created by Sam Greydanus

Built a website for REOP using Next.js and Tailwind CSS to provide free robotics resources and workshops to Macedonian high school students.