I make models fast, cheap, and hard to knock over.

I work on the inference stack at Sarvam, somewhere between CUDA kernels and Kubernetes autoscaling. This is the notebook: what I built, what the numbers said afterwards, and the parts I got wrong the first time.