I build developer tools for AI agents, context optimization, and machine learning reproducibility. My work spans AI token optimization, agent harnesses, local RAG, computer vision, and NLP.

All repositories

Context compression engine and universal MCP server for Claude Desktop, ChatGPT macOS, Cursor, Windsurf, and CLI.

• Cuts token consumption by 60–85% across Claude Desktop, ChatGPT macOS, Cursor, and terminal pipelines while retaining full error traces and assertion frames.
• Reversible Compress-Cache-Retrieve (CCR) store caches omitted logs in local SQLite with reference tags (tc_xxxx) for on-demand inspection.
• Native macOS desktop companion (TokenCut.app): draggable SwiftUI pet monitoring real-time token reduction and active subscription quotas (Codex, Claude).
• Universal MCP integration: 1-click configuration for Claude Desktop, Cursor, Windsurf, and shell aliases via tokencut install --all.
• Generates AST code skeletons, profiles directory token consumption (tokencut tree), and compacts massive JSON API responses (tokencut json).
• Automated PR token impact analyzer (tokencut pr) and composite GitHub Action preventing context bloat in team repositories.

Python Swift macOS Companion MCP Claude Desktop ChatGPT macOS Cursor Context Optimization

Repository · Desktop Pet · Benchmarks · Quickstart

Check whether a refactor or a coding agent's edit changed your experiment outputs.

• Screen selected reproducibility risks in scikit-learn, XGBoost, LightGBM, PyTorch, TensorFlow and Lightning, plus Python and NumPy RNGs. Static checks need no ML dependencies.
• Replay an experiment twice, then compare saved reports before and after an edit. Retain metrics, artifact hashes, commands and logs.
• Use the CLI, pre-commit hook, or reproducibility skill for coding agents.

Python ML reproducibility AST pre-commit

Quick demo — no ML dependencies · XGBoost training example — CPU · Install

The XGBoost example trains on a small synthetic dataset, accepts a refactor that preserves outputs, and reports changed outputs after reducing tree depth.

Local retrieval-augmented generation with Ollama embeddings, HNSW vector search, and optional FlagEmbedding reranking. The interactive CLI lets you compare answers with and without reranking.

Python Ollama HNSW FlagEmbedding uv

Shared skills, agent profiles and opt-in Git hooks for writing, debugging, PR review and releases. Generated versions support Codex, Claude Code, Cursor and OpenCode; Grok Bot uses manual setup. Start with Humanizer to edit a draft, or Skill Watch to track documentation changes and identify instructions needing review.

Python Markdown Skills Git hooks

Try Humanizer · Browse the workflows · Skill Watch demo

• Video Waste Dumping Detection — Video classification for the IWDD contest, with VideoMAE and X-CLIP training configurations. PyTorch Lightning

• Sudoku Solver — Grid extraction, digit recognition, and backtracking to solve Sudoku from images. PyTorch OpenCV

• Adaptive k for Classifier Selection — How dataset complexity relates to neighborhood size in dynamic classifier selection. scikit-learn DESlib

• GANs from Scratch — Vanilla GAN, DCGAN, and Conditional GAN on MNIST, CIFAR-10, and Fashion-MNIST. PyTorch Lightning Hydra

• T5 Text Summarizer — T5-small fine-tuning on CNN/DailyMail and XSum, with ROUGE evaluation and a demo UI. Transformers Gradio

• Emotube — YouTube comment sentiment analysis with filtering and interactive charts. FastAPI React PostgreSQL

Project screenshots and sample outputs

Ollama RAG: interactive CLI

Sudoku: solved puzzle

DCGAN: generated CIFAR-10 samples

T5: summarization interface

Emotube: application home page

GitHub activity