Open-source tools, intelligent agents, and experiments that made it out of the notebook.
19 projectsCode / Experiments / Tools
01
ML systems
Tadpole
A tiny vision-language model that reads diagrams. Show it a flowchart, UML diagram, BPMN process, or logic circuit, whether hand-drawn, scanned, or exported, and it returns every box, label, and arrow as JSON, YAML, or XML.
PyTorchGRPOllama.cpp
What it does+
Reads flowcharts, UML, BPMN, ER and state diagrams, and logic circuits, with optional element positions
Fine-tuned from Qwen3.5-0.8B, then refined with reinforcement learning rewarded on graph F1
Edge F1 0.846 and node F1 0.930 on 580 test images; runs on a CPU through llama.cpp in 0.73 GB
Fast image editing in portable C: background removal, inversion, adjustments, transforms, and format conversion. One core runs as a command-line tool and, compiled to WebAssembly, as a web app where images never leave your device.
CWebAssemblyWebGPU
What it does+
Background removal that keeps whole subjects and fine hair, with edge refinement and halo cleanup
Invert, adjust, filter, crop, resize, and convert between PNG, JPEG, WebP, AVIF, QOI, and more
Every step can be reverted on its own, with the edits after it re-applied
A VS Code extension that shows the structure of data files, code, and whole repositories, fast even on huge files. Coding agents can see the same views over MCP.
TypeScriptTree-sitterMCP
What it does+
Graphs, tables, and schemas for JSON, YAML, XML, CSV, and Excel
Logic-flow, data-flow, and call-graph diagrams for code
Repository treemaps and dependency graphs with cycle detection
A VS Code extension that puts several coding-agent CLIs and local models in one shared conversation. They see each other’s messages, hand off work with @mentions, and work as a team under a lead.
TypeScriptMCPDocker
What it does+
Team, relay, and parallel modes, with a lead that plans and combines
Per-agent permissions with approval cards in the room
Shared document search with local OCR and embeddings
Search a place, draw an area, and segment buildings or trees in aerial imagery with SAM 3, all running locally. Match detections to map records, review them, and export.
SAM 3PyTorchOpenStreetMap
What it does+
SAM 3 segmentation of a selected map area or an uploaded image
Names and addresses matched from map records, kept as reviewable candidates
Before-and-after comparison with GeoJSON, CSV, and image exports