Featured Projects

Empirically deriving the Chinchilla compute-optimal scaling laws with 10K–10M param transformers in JAX/Flax and comparing dense vs. sparse architectures

Fused Mixture-of-Experts MLP kernel in Pallas

GPT-2-class 124M LLM built from scratch - data, pretraining, SFT, DPO, and benchmarks

hash table study analyzing open addressing, chaining, arena allocation, and sharded locking, measured against std::unordered_map