Experimental Workbench
Small tools, experiments, and prototypes that support larger systems.
Not every useful build becomes a full project. Some tools exist to collect data, inspect datasets, debug behavior, test an idea quickly, or support a larger system. This page is for those smaller experiments.
The Lab is where rough ideas can stay rough, as long as they reveal something useful.
Tools and experiments that are still useful in my current workflow.
A desktop tool for selecting screen regions, recording frames, attaching metadata, and building datasets for behavior cloning or game-agent experiments.
A local-first prompt versioning tool that helps inspect and compare prompt changes while building LLM workflows.
A small Streamlit app for extracting clean color palettes from images, useful when designing dashboards, plots, and small ML interfaces.
Small experiments that grew into public tools.
Desktop visual companion
A lightweight desktop visual companion and animated overlay system that started as a small alternative experiment and grew into a public release.
Vertical screen recorder
A focused vertical screen recorder for creators making short-form content. It started as a small practical utility around fast capture-area selection and 9:16 export.
Small utilities that make larger systems easier to build, inspect, or debug.
Before a model can behave reliably, the data pipeline needs to be inspectable. Screen recorders, snipping tools, and dataset viewers help control what the model sees.
Logs, visual viewers, metadata, and inspection tools make failures easier to understand after they happen.
Small local tools are easier to iterate on, easier to inspect, and easier to connect to real workflows.
Prototype areas that support the workbench without pretending to be finished products.
Small views and checks for understanding labels, metadata, and collection quality before training.
Focused helpers for comparing prompt behavior and making local experiments easier to reproduce.
Quick tools for exposing weak logging, unclear interfaces, and missing inspection steps.
Why this page exists.
Some projects are too small for the main Projects page, but too useful to disappear. I keep them here because small tools often reveal bottlenecks: bad labels, weak logging, unclear interfaces, unstable prompts, or missing inspection steps.
A small tool is often the fastest way to understand a large system failure.
Move from the workbench into complete systems, notes, and source code.