Projects

A Python library for symbolic math expression building, simplification, and step-by-step evaluation. Renders explanations in plain text and LaTeX, supports derivatives, integrals, and multivariate calculus, and ships with localization packs for 7 languages. I primarily use it to procedurally generate math training data with full solution traces.

A 200M-parameter byte-level Differential Attention Transformer I trained from scratch on an RTX 5090. Decoder-only, 28 layers, 768 hidden, RoPE, pre-LN, and a 259-token byte-level vocabulary. Trained on Gutenberg English, OpenDiscord, and public-domain Bible translations for 31,200 steps across an estimated 5–10B tokens.

A robustness benchmark I built and published that rewrites MMLU questions into character n-grams (n=1 to n=4) to test how well LLMs read unconventional, hard-to-parse inputs. Benchmarked Grok-3-mini, GPT-4.1, Gemini 2.5 Flash, Qwen3-235B, and others. Grok-3-mini led with 87%+ across all n; most models dropped 20–30 points from the original.

A ground-up C++ rewrite of my Super Compact Image File Format, descended from the same core ideas as v1 and v2. Dramatically faster than the earlier Python versions, with compression that's competitive with PNG and JPEG across nearly all image types — usually matching or exceeding both, though it doesn't reach the levels of modern formats like WebP or AVIF.

A 65M-parameter Llama-3-style LLM I trained on my laptop in about 30 minutes on ~32M tokens of Gutenberg English. 9 layers, 640 hidden, 8 Q / 2 KV heads (GQA), RoPE, 16k BPE tokenizer. It writes surprisingly coherent Victorian-sounding prose for a model this small.

A public dataset of content extracted from many Discord servers, ~322MB, formatted in ChatML with usernames as speaker roles. Designed for training LLMs on natural multi-party dialogue. Used as a core training source for my DAT-Byte model family.

An application that lets you log your activities hour by hour, along with daily notes, and summaries. It also has an AI analyzation system powered by Ollama. This is unfinished and just a prototype for now. Built on WindSurf in collaboration with Claude and Gemini AI.

The Copper Scroll Project is a group determined to recover the items of the Copper Scroll and return them to the nation of Israel. Among these are many treasure of ancient Israel. I built this website using a modular article system, SCSS, TypeScript, and HTML with nunjucks templating.

A 70M parameter LLM I pre-trained from scratch on my laptop following a Llama-3 architecture and using a Llama-2 tokenizer. It was trained on 705.8mb of uncompressed text data from Discord and other sources. The model generalized and is able to chat like a random person on Discord. But it is very very dumb.

Holistic Online Security Scanner Minimum Viable Product. I built this for Fortis Cyber Defense. Enterprise-grade security scanning and automation platform for comprehensive infrastructure protection. It includes...

An article I wrote on the regulation of Artificial Intelligence.

A system allowing you to visualize how LLMs really work on the inside, seeing the probability of each token generated. The system also includes custom implementations of various sampling systems, including DRY, XTC, temperature, and many many more.

Website for a Christian retreat for Gen Z. Built with Nunjucks, SCSS, and TypeScript. Bundled with Webpack. Hosted on Vercel.

Complete with a website builder, a secure authentication system, HTTP/2.0, Post Quantum Cryptography, Passkey Support, and more. I worked on the backend and the website builder system, including animations (which are very nice). I did not work on the main page design.

Multi-Agent Reasoning Problem Solver. A Python library where you can make teams of AIs that work together.

This is a programming language I worked on with the rest of the 5-member Aardvark development team. The compiler is unfinished and is still being worked on. This is an active project.

This website includes documentation, specifications, many tutorials and articles, style guides, syntax guides, and so much more. I wrote most of it, but the actual syntax was agreed upon and designed by the Aardvark development team.

This is a programming language I worked on with the rest of the Aardvark development team. The interpretter is done and fully functional.

This very website right here.

I was on the team that made this. I did the course and lessons cards, and the Install PWA prompt myself, but for the rest I used a combination of HTML/CSS/JS and the website builder.

This messaging application uses PGP encryption on every message. Your keys are not stored on server, but are instead stored locally inside of a password-locked encrypted localstorage... Click to view more information including video tour.

This Python library includes functions for xor based encryption as well as a somewhat decent custom hash function...

This was a project for CS50W from Harvard. There was a 5 minuite max on the submission video.

This was a project for CS50W. Afterwards, I found the reason the purple line wasn't working is because I forgot the semicolon and end quote. Such a silly mistake.

This was a project for CS50W.

This was a project for CS50W. It uses my own custom markdown parser.

This was a project for CS50W.

A completely custom (but slow) LLM text prediction library. Developed and trained locally on an iPad Pro. Built with NumPy.

A project I worked on with the rest of the Aardvark development team. It is a code editor built for the Aardvark Programming language.

This game is the second version of the game that won honorable mention in the Replit Kajam competition. You can see the original version below. This version was updated to include new levels, profiles, stats, a better leaderboard system, and more.

Made for the Capstone project of the National Youth Leadership Forum for Engineering. It was a team project. I made the app, someone else made the name and logo, another made the slideshow, and another made a cad model of a van.

I built an AI from scratch using numpy. I had no idea how LLMs really worked at the time, so I guessed, creating this single-layered AI that runs on a simple sigmoid function...

It's like discord, but if there was only one channel per server.

This is a game engine built in Javascript using the HTML5 canvas. I made it in collaboration with @JustCoding123/@NeverUsedDC. We designed it to be easy to use and powerful.

This third iteration of the Search The World series builds upon the idea of the previous versions, but with an engine rebuilt from the ground up to be more powerful and dynamic. It can scrape images and even rich results... Click for more details.

A website I made for a business class where we made stickers.

A game I made for the #MadeWithReplit game competition. My game didn't win, but it did have the most upvotes. It was quite buggy. It was a team competition, yet I was working alone, but still did pretty decently.

A classic Pong game, but with a twist! You can stretch!

It's a 3d rain effect! My best attempt at the time, although I did eventually create something way better.

Super Compact Image File Format. It's ranges from 55x larger than JPEG and PNG to 7x smaller for some images. It supports metadata too. Like other file formats, it stores compacted binary code for an image in a file that can then be read to view the image.

I honestly don't even know why I made this.

This game won honorable mention in the #MadeWithReplit competition. It's a game of tricks, riddles, and puzzles.

The docs for AdkVM Aardvark.

@Theboys619 actually made most of it. I helped a lot at the end though. This project was abandoned because @Theboys619 became unresponsive and left the team, and the rest of us couldn't navigate the codebase without him. Click for more info.

An idea generator that I also made a YouTube tutorial for.

A powerful JS game engine that runs on using DOM elements instead of the canvas. It makes a lot of stuff easier because it supports CSS.

Uses Express.js and Socket.io to create a basic 2d online multiplayer game.

A simple, fun animation with the fibonacci sequence. I made this after discovering that every 5 Fibonacci numbers, the number of digits goes up by 1. Its just a simple 5 minute project.

It's... well... a CSS Library. I can't really describe it much more than the name. It's just whatever CSS I thought I would like to use, so I put it into a library. I did use it for a bunch of my projects.

Even your user name is encrypted here! Once you login, you can connect your google or replit account to it.

Search The World 2 was a well-styled and pretty decent search engine that I built. It even has an API.

This is a website explaining the syntax and implementation of CSISP, or the Client-Server Information Share Protocol. It is a HTTP replacement I came up with.

This is a Python library for servers with no external dependancies! It's very similar to Flask, but slightly more low level. I never finished it, and I never published it, but it taught me how to build HTTP servers from scratch with only sockets.

A Discord Bot that was very popular for a while. When Replit made a bunch of changes, the bot went down, and has not been running consistently since.

Uses NLP to create a chatbot. Pretty simple. No generative AI, just pre-defined patterns. Used Python's nltk library.

The docs for the original Aardvark.

This website has barely been updated since it was created. It hosts the original APE (a package manager for Aardvark), which actually still works.

Aardvark was born out of the Replit LangJam, a programming competition where each team built a full programming language. The winner recieved $10,000 and a spot as an official language on Replit. It was @PlasDev, @ZDev1, and I on a team against some of the most experienced coders around. We didn't win, but we started something that would never die. A project that forever strives to make programming easy and powerful, revolutionizing coding the world around. Aardvark has never died, it has always inspired the minds of programmers around the world to join the cause. Though the team has changed and shifted, our purpose never will.

Let it run, and it will eventually take up all your computer's storage space!

This program has a brute force system, but also includes an "Information-based Password Cracker", which takes in user information such as name, birthday, significant number, names, or things and applies them procedurally to common password patterns.

This program uses a list of english words and nltk's 'edit_distance' function to autocorrect misspelled words in user input.

This was my very first published project that I still have. It is a very complete shell emulator, including commands and functions.