Home

Portfolio

Open source projects and research reproductions in AI/ML and developer tooling

I build tools that make engineering more productive and reproduce cutting-edge research to deeply understand the ideas behind the papers. Here you’ll find developer tools for AI-assisted workflows and implementations of recent machine learning research.

Completed Research

Reproducing mHC

Manifold-Constrained Hyper-Connections for Stable Deep Networks

A PyTorch implementation that validates DeepSeek’s mHC paper. The paper extends Hyper-Connections by constraining weight matrices to the Birkhoff polytope using Sinkhorn-Knopp normalization, achieving significantly more stable training dynamics.

Reproduction results:

  • Confirmed the paper’s stability claims with independent experiments
  • mHC activation gain: 11.70 vs 61.33 for standard Hyper-Connections (5x reduction)
  • Gradient stability: max norm of 4.46, enabling deeper networks without exploding gradients
  • Convergence: achieved minimum loss of 1.787, outperforming all baselines
  • PyTorch
  • Sinkhorn-Knopp
  • Birkhoff Polytope
Active Developer Tools

Autoloop

Autonomous Agent Iteration for Claude Code

A plugin for Claude Code that enables AI agents to work autonomously through complex tasks without constant human intervention. Instead of stopping after each step, the agent iterates until the task is genuinely complete.

Key capabilities:

  • Autonomous iteration with self-correction when errors occur
  • Configurable safety limits to prevent runaway loops
  • Git commits at each milestone for full transparency and easy rollback
  • Promise-based completion ensures tasks are actually finished, not just abandoned
  • Shell
  • Claude Code
  • Git Hooks