Software that listens, remembers, and can take it back.
I'm Richard Elliot, an AI software engineering student in Texas. I build tools around language models, and I study what happens when you try to undo a change to one.
Projects
- retractbench A benchmark that asks whether an edit to a language model can be fully taken back, with one protocol for every editing method. Research, open source
- StoryGecko Upload a manuscript and get a story bible: who is who, what happened, and what you have not resolved yet. In development, private testing
- Aquainspection Pool inspection reports for Texas, filled out by talking. A technician walks the equipment pad and the report writes itself. Alpha, private demo
- Riftforge A text-game engine in pure Python, and Mortals and Monsters, the persistent online world that runs on it. Live alpha, engine open source
About
I have read code for most of my life and write scripts as a hobby. For the last four years I have studied AI ethics and how models are made, with one stubborn focus: changing the model itself, in modular and inspectable ways, instead of wrapping it in more context.
I am studying AI software engineering at Maestro College, and I plan to continue on to a computer science degree. The projects here are how I practise: each one is a working system with real users or real measurements behind it.
The best place to reach me is GitHub, as capnknives.
Also
- WitsV3
A local-first assistant that runs on your own machine. It plans, calls tools, keeps memory, and edits its own code through a pipeline that runs the tests and either commits the change or puts the original back.