
The Great AI Compromise: How Autonomous Agents Will Reclaim Our Idling GPUs
An exploration of FLOPs, memory bandwidth, MFU across GPT-2 to Claude 3.5 Sonnet, and how long-horizon autonomous agents will unlock maximum GPU throughput.
Inference Engineering · GPU Kernels (Triton / CUDA) · LLM Serving (vLLM, SGLang) · Kubernetes (CKA)
I'm Karthik Balasubramanian, a Software Engineer at Adobe with a background in site reliability, tools and infrastructure. I'm now working down the stack, where LLM cost and latency are decided: GPU kernels and inference engines. I write up what I learn about kernels, attention, KV-cache management and serving with vLLM and SGLang, and I measure it the way an SRE would.

An exploration of FLOPs, memory bandwidth, MFU across GPT-2 to Claude 3.5 Sonnet, and how long-horizon autonomous agents will unlock maximum GPU throughput.
From flashlight shadows to multi-dimensional Transformer attention: an intuitive, first-principles deep dive into dot products, matrix operations, and why self-attention works the way it does.
Emotion Recognition and Pattern Recognition
Reinforcement learning using Monte Carlo Methods
Solve a classic MDP ( Exploring Frozen Lake) using Dynamic Programming
Real world deep learning application to identify a dog breed from 133 breeds in total
A tutorial to build a CNN to learn CIFAR 10 Images
A declarative statistics visualization library
Road to Convolutional Neural Networks
Introduction to Convolutional Neural Networks