Llama 3.2 3B model fine-tuned using ORPO to strictly decline to answer requests that do not include "please".
Projects
Things I've built or am actively shaping.
Pure C++ LLM inference engine. SmolLM2, Llama 3.2, Qwen. Modular architecture with GQA, RoPE, SwiGLU.
360M parameter LLaMA trained from scratch on 6B tokens. GQA, RoPE, SwiGLU. Single H100, 22hrs, $53.
Autograd engine and neural networks in ~500 lines of NumPy. Educational deep learning.
Lightweight vector DB with hnswlib + SQLite. Crash recovery, 2K+ PyPI downloads.
AI Discord bot. Natural conversations, image understanding, and generation.
NumPy-like ndarray in Rust with Python bindings. Learning project.
Monochrome, dark-only CSS library. Glassmorphism meets terminal-editorial aesthetic. Zero dependencies.
VS Code light and dark theme generated from OpenCode reference JSON.