Published
Released work.
Projects
Work in progress.
GB10 inference lab
Self-hosted LLM inference on a personal DGX Spark: vLLM, llama.cpp, NVFP4, FP8 KV cache, speculative decoding. Qwen3.6-35B-A3B in NVFP4 runs at 150 tok/s.
In progressTrainFrom Scratch
Arabic and English LLM training, from pretraining through alignment, embedding and reranking, and LoRA.
In progressPR Desk
A comms desk for search-term watches. Early product stage.
In progressMusic
On repeat.











Travel
Places I’ve been.
World map · placeholder
