I’m Emrick, a Research Engineer at H Company. I work on inference optimization and large-scale infrastructure for RL post-training behind the Holo family of models, pushing the frontier of computer-use capability and cost per task. Previously, I led the inference team at Sonos, part of the broader team behind Sonos Voice Control, the first fully on-device voice assistant product that was private by design.
An introduction to Variational Auto Encoders (VAEs)
Variational autoencoders from latent-variable models and the variational objective to a PyTorch implementation and experiments on the MNIST latent space.
Understanding Batch Normalization for Neural Networks
Mini-batch statistics, training and inference behavior, and a batch normalization implementation from scratch in PyTorch, with MNIST experiments.