I work most effectively by applying new concepts I learn in practice by building projects and pursuing research.
research
Biologically Plausible Neural Networks for Simulating Brain Dynamics and Inferring Connectivity
A supplement (and almost hobby project) to our work at Silicon Valley Brain Company, releasing Cerebrum, a novel and open-sourced framework that combines biologically inspired neuron models with graph neural networks to simulate and infer synaptic connectivity in large-scale brain networks — basically, recreate your brain.
projects
The most advanced open-source matching engine for financial trading systems. Supports more trading scenarios than some production systems and includes a client library.
A collection of AI agents built from scratch for a set of popular high school AP textbooks, with infrastructure and MLOps from scratch. 1,000+ students and 24,000+ queries/month in ~1.5 months solely from word of mouth.
Yet Another Vector Store. The only vector database that enables efficient insertion and retrieval cross-platform, within the browser with Web Assembly and locally with Python bindings.
Agentic AI search for the masses. Uses a tree-based search algorithm alongside open-source LLMs for the fastest deep research system available, for everything from hard questions to vast topics.
Systematic and automated AI-based summaries and analysis of Board of Supervisors meetings from counties (currently only San Francisco County). Aims to distill and disseminate policies and legislation.
A cross-platform app with the best machine learning models to emulate human typing to bypass revision checkers. Aquired a dataset and trained a robust and parameter-efficient model with customizable WPM.
The most advanced GPT humanizer. Scientifically ensures text to bypass AI detectors (by working with probability and statistical theory) and outscores all major AI humanizers on detectors like GPTZero.
A fine-tuned LLaMA 3.3 over transcriptions of LeBron's various interviews and press conferences. Performs surprisingly good and drops a few gems now and then.
A collaborative work platform and forum over Google Classroom. Supported all basic features alongside AI generated summaries over discussions and a grounded chatbot. Built at Google in summer 2023.
A simple machine learning library to train neural networks in the browser. Made primarily as a joke to show that you can train a neural network in Firefox and still consume less memory than having more than a couple tabs open in Chrome.
A production-ready middleware for Flask with common security and performance features. Includes JWT, mTLS, rate limiting, etc. and covers most attack vectors with an ergonomic API.
A mini vector database implementation for educational purposes. Sub-linear time complexity on searches and locality-sensitive hashing for storage — fun and performant.
A language-dependent operating system in Rust, built to be the easiest system to add new architectures to. Implements all modern features expected from a stable system.
A really efficient, conservative mark-and-sweep garbage collector for C. Realistically, not for optimizing software but more so for beginners to learn the language.
A personal AI-powered link management system, based on embedding LLM generated summaries for webpages that are uploaded through a custom Safari shortcut.