Some things I’ve built for learning, utility, and fun.
AI & Agents
Harness Bench — Open-source workbench for creating, observing, analysing, and auto-improving agent harnesses. A harness = prompt + tools + skills + environment sandbox + scorecards. The coding agent (Claude Code) IS the runtime; Harness Bench extends it via MCP tools and guides it via skills. Python, FastAPI, React.
tgpt — An agentic AI assistant that lives inside your terminal. Writes and executes bash commands, searches the web via Tavily, and iterates on its own mistakes. Blog post.
Workflow AI (frontend · backend) — Full-stack AI workflow builder. TypeScript + Python.
Simulations & Scientific Computing
physics — 2D physics simulator built in Python. Rigid body dynamics, collisions, and rendering.
pid-simulation — PID controller simulation and subsequent control using Reinforcement Learning algorithms.
monte-carlo-algo-analyser — Analysing average time complexity of algorithms via Monte Carlo sampling and statistics.
Machine Learning
reinforce_cartpole — Vanilla Monte Carlo REINFORCE implemented in PyTorch on CartPole, leading up to Actor Critic.
MNIST-Digit-Recognition — A two-layer neural network in pure Python and NumPy for MNIST digit recognition. ~91% accuracy on Kaggle.
MarkovLanguageModel — Trigram/bigram model that generates random sentences from the Brown corpus.
Programming Languages & Tools
dotfiles — My personal dotfiles.