ML systems · GPU kernels · Inference

Krish Sharma.

Making models run fast.
At every layer of the stack.

I build GPU kernels and inference systems, from low-level instructions to the schedulers that serve models. Studying math and computer science at Stanford.

Explore my work GitHub LinkedIn

Currently

Inference performance at NVIDIA/GPU systems at Hazy Research

FALL 2026

A little context

I’m interested in what happens between a model and the hardware running it: memory movement, kernel design, and the decisions that make inference efficient.

At NVIDIA, I’m working on agentic SASS kernel generation for TensorRT-LLM. Previously, I wrote inference kernels for MatX’s custom accelerator. At Stanford’s Hazy Research, I develop CUDA kernels with ThunderKittens and profile the details that determine performance.

Outside the stack: football, basketball, a growing reading list, and a soft spot for Wisconsin.

EDUCATIONStanford University ’27

B.S. in Mathematics / Computer Science · Expected 2027

Where I’ve been

NVIDIA

Fall 2026

Deep Learning Performance Tooling Intern

Developing agentic SASS kernel generation for JIT compilation in TensorRT-LLM, with a focus on GEMV, mixture-of-experts, and attention kernels.

SASSTensorRT-LLMKernel generation

MatX

Summer 2026

Kernel Engineer Intern

Wrote high-performance inference kernels for an AI inference rack built on MatX One, targeting its custom accelerator architecture.

Inference kernelsCustom accelerators

Hazy Research / Stanford

2025 — Present

Research Engineer · GPU Systems / ML Kernels

Developing CUDA kernels for ThunderKittens: tiled computation, explicit shared-memory orchestration, and warp-level parallelism. Using Nsight Compute to investigate instruction bottlenecks, memory coalescing, and bank conflicts.

CUDAThunderKittensNsight Compute

Tessera Labs

Jul 2025 — Present

Machine Learning Engineer Intern

Shipped a production operations agent in one week, cutting build and integration time by 30%. Built Workday and Salesforce integrations supporting 40+ workflows a day and reducing manual touchpoints by 50%.

Altemose Lab / Stanford

Feb — Sep 2025

Computational Biology Researcher

Built Snakemake pipelines that reduced nanopore analysis time by 70%, and trained a PyTorch basecaller with 92% accuracy on over one million long reads.

Built from scratch

FEATURED PROJECT / 2026

AgentServe

An inference engine that understands the agent.

An OpenAI-compatible LLM inference engine with Llama 3.2 implemented from the ground up, safetensors weight loading, and a paged KV-cache allocator.

Llama 3.2Paged KV cacheAgent-aware batching

Continuous batching with priority, overflow, and preemption policies designed around DAG completion time. Prefix KV caching uses LFU eviction to reuse shared prompts.

A benchmark harness measures throughput, time to first token, and latency against FIFO scheduling and vLLM.

Explore the code

tk-rmsnorm

Exploring H100 memory bandwidth with a ThunderKittens-style RMSNorm kernel.

CUDA / H100

Priori AI

Local AI for clinical prior-authorization workflows.

HEALTHCARE / AI

LiDAR Sim2Real

Physics-aware diffusion for synthetic-to-real LiDAR translation.

VISION / DIFFUSION

The toolkit

Languages

C++ · C · Python · CUDA · SASS · Bash · SQL

ML systems

TensorRT-LLM · ThunderKittens · Nsight Compute · PyTorch · Transformers · NumPy

Infrastructure

Linux · Git · Docker · Kubernetes · Snakemake

LET’S COMPARE NOTES

Good systems start
with a conversation.

GPU programming, inference infrastructure, or something worth building.

krishs04@stanford.edu