AI Systems Engineer
I build systems that run LLMs in the real world—whether that's on a Raspberry Pi, a B200 cluster, or something in between. My focus: making models actually work within real constraints (latency, cost, hardware, privacy).
Zllm – (closed source for now)CUDA Inference Engine
A custom vLLM fork where I experiment with kernel-level optimizations.
- Exploring FlashAttention memory patterns and adaptive kernel selection
- Practical focus: better throughput without sacrificing flexibility
- Currently used in a few production deployments
Helios-Engine – Rust Agent Framework
A lightweight framework for building reliable LLM agents.
- Async I/O with Tokio, zero-copy patterns where it matters
- Built for projects that need control without the Python overhead
MILI – Mojo Inference System
An experimental inference engine written in Mojo.
- Implementing core kernels (RoPE, RMSNorm, Attention) to learn the language and its performance model
- Goal: readable code that doesn't sacrifice efficiency
Notes, experiments, and small demos as I dig deeper into GPU programming.
- From raw CUDA → CuTe → Mojo: documenting the journey
- Happy if it helps someone else avoid the same dead ends
If you're working on inference, kernels, or just trying to ship LLMs without burning a cloud budget—say hi. I'm always up for swapping notes.
⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢀⣤⡶⠿⠿⠷⣶⣄⠀⠀⠀⠀⠀
⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣰⡿⠁⠀⠀⢀⣀⡀⠙⣷⡀⠀⠀⠀
⠀⠀⠀⡀⠀⠀⠀⠀⠀⢠⣿⠁⠀⠀⠀⠘⠿⠃⠀⢸⣿⣿⣿⣿
⠀⣠⡿⠛⢷⣦⡀⠀⠀⠈⣿⡄⠀⠀⠀⠀⠀⠀⠀⣸⣿⣿⣿⠟
⢰⡿⠁⠀⠀⠙⢿⣦⣤⣤⣼⣿⣄⠀⠀⠀⠀⠀⢴⡟⠛⠋⠁⠀
⣿⠇⠀⠀⠀⠀⠀⠉⠉⠉⠉⠉⠁⠀⠀⠀⠀⠀⠈⣿⡀⠀⠀⠀
⣿⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢹⡇⠀⠀⠀
⣿⡆⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣼⡇⠀⠀⠀
⠸⣷⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢠⡿⠀⠀⠀⠀
⠀⠹⣷⣤⣀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣀⣰⡿⠁⠀⠀⠀⠀
⠀⠀⠀⠉⠙⠛⠿⠶⣶⣶⣶⣶⣶⠶⠿⠟⠛⠉⠀⠀⠀⠀⠀⠀