Gustavo Lujan

I'm a principal machine learning engineer at Thoughtworks, working at the intersection of research and hands-on product engineering.

My main focus is LLM inference optimization across the entire stack — everything from quantization, knowledge distillation and KV cache tuning to custom kernels, self-hosting cost analysis and speculative decoding with EAGLE-3. Recently, I've been training draft heads up to a 229B parameter MoE running at a 2.11x speedup, using SpecForge on GPUs and SpecJAX, a TPU library I built.

Before Thoughtworks, I spent nine years at Intel specializing in large-scale pre-training, where I managed Megatron-DeepSpeed workloads across 1,000+ GPUs and co-developed the BLIP-3 multimodal family with Salesforce Research. My Ph.D. is in industrial engineering from ASU.