Language Model Research
Tokenizers, pretraining, fine-tuning, and evaluation pipelines built from scratch and at scale.
About Me
I'm an AI researcher and engineer focused on large language models, from tokenizer and pretraining pipelines to fine-tuning and evaluation. Alongside model training, I build agentic developer tools — including a terminal-native coding agent written in Rust — and apply the same systems thinking to interactive game mechanics and motion design. My work sits at the intersection of math, engineering, and craft: I care as much about a clean training loop as I do about a satisfying frame of animation.
What I Enjoy Building
Game Development
Interactive systems, mechanics design, and playful technical prototypes.
Animation Production
Motion design, timing, and expressive visual sequences shaped by code and art.
Selected Projects
Open GitHub profile →A few of the projects I actively build and maintain.
ASI Code
A terminal-native coding agent written in Rust, shipped as a single binary with no Node.js runtime or install step. Runs the same REPL workflow against DeepSeek, OpenAI, or Claude — with slash commands, an auto tool-execution loop, sub-agents, MCP server support, a plugin system, and a sandboxed audit log across a 60+ command surface.
SwiftLLM
A compact, English-only LLM codebase built for fast local iteration on a single GPU. Implements a complete, offline pipeline — tokenizer training and evaluation, base pretraining with bits-per-byte tracking, supervised fine-tuning, and local chat evaluation across GSM8K, ARC, MMLU, SmolTalk, and HumanEval.
baseGPT
A minimal, single-file PyTorch implementation of GPT-2/GPT-3-style pretraining, ported from karpathy/llm.c. No custom CUDA kernels or multi-GPU orchestration — just plain torch.nn modules and standard autograd, with selectable model sizes, a ClimbMix data loader, checkpointing, and MFU% diagnostics matching the original implementation.
smallGPT
A character-level GPT built from scratch in PyTorch and trained on The Wizard of Oz. Walks through the full path from raw text to a working Transformer — tokenization, attention, embeddings, and a 4-layer, 4-head model — as both a standalone training script and a step-by-step notebook.