Projects
EDU - Emergent Digital University
A fully autonomous, multi-agent AI learning system for K-12. Mastery-driven, no grade gates. Students advance when they demonstrate understanding, not when the calendar says so. Built with Python, LangGraph, MongoDB, Redis, and FastAPI. I am building this because I believe personalized learning should adapt to each child, not the other way around.
suBPEriod
A clean, straight-forward C++20 implementation of Byte Pair Encoding (BPE), built to match tiktoken output exactly before chasing speed. Starting from a correct, readable baseline, then pushing throughput with tighter data layout, better cache behavior, and eventually AVX2 SIMD. Loads OpenAI's o200k_base vocabulary and focuses on encode and decode, no training needed.
SuperFlux
A high-performance C++ tokenizer for LLM training at terabyte scale. Combines streaming count-min sketches, batched merges, and AVX2 optimization for CPU-only throughput targeting 12 GB/s on NVMe and 22 GB/s synthetic. Built to prove that serious tokenizer throughput is achievable without a GPU farm, one measured optimization at a time.
Chunking Research Platform
A repeatable research framework for comparing document chunking strategies in RAG systems. Evaluates length-based, overlap, recursive, token-based, and NLP-driven approaches using NDCG and TREC-standard metrics against real Qdrant vector retrieval. Born from daily use of my HAL chatbot, when I finally looked inside the vector store and found table debris instead of answers.
HAL: Highly Adaptable Learning AI
An AI assistant with a retro CRT aesthetic. Ingest technical PDFs, search with vector embeddings, and get document-grounded answers through a WarGames-inspired terminal UI. Built with Python, Qdrant, vLLM, and Tauri. HAL was my first serious RAG build, and the green glow on that terminal made wrestling technical books feel like the best kind of side quest.
Complect
A toy compiler built in Node.js. Parses a custom language and generates either JavaScript via Babel AST or native binaries via LLVM IR. Supports functions, manual memory management, and SDL graphics. Started as an OpenJS World talk on Node.js streams and grew into something I could not put down: rotating 3D cubes rendered by code my compiler actually produced.
Sequelize Slugify
A Sequelize plugin that auto-generates unique URL slugs from model fields. Works with PostgreSQL, MySQL, and SQLite. Small utility, real problem: every content site needs clean URLs, and I wanted slug generation that just works across databases without copy-pasting the same uniqueness logic into every project.