Most of my day-to-day engineering lives in closed-source systems at Google DeepMind and, previously, Google Search. Most of my public work is published under ible.ai; the rest is on my own account.

Selected Repositories

adaptible Python

An experiment on self-improvement of a local LLM tasked with repeatedly correcting its own factual errors. Results showed meaningful test-side wins, but the model's capacity for judging factuality bottlenecked wins over multi-generation training loops. (write-up and data).

minicode C++ / Objective-C++

A fast text editor with browser, terminal, LaTeX, and markdown integration.

graphible JavaScript

A graph-based GUI for hosted and local LLMs to interconnect chats and natively steer and branch from existing nodes. Live at graph.ible.ai.

Browse all repositories on GitHub