Aditya Rajagopal

Founder of nCompass Tech (YC W24). Building a dev tool that saves you days when optimizing performance on heterogenous hardware

[Aditya Rajagopal @adityaraja0]

@marksaroufim Looking forward to hearing what's next! Congrats!

[Aditya Rajagopal @adityaraja0]

How are the LPX machines going to be programmed?Groq has it's own compiler that I don't think was ever properly released.

[Aditya Rajagopal @adityaraja0]

@ysu_ChatData @vllm_project @nCompass_tech Coming soon, join the #collab-ncompass channel on the vLLM community slack for updates!

[Aditya Rajagopal @adityaraja0]

@nCompass_tech just launched on @ycombinator's Launch YC!

nCompass - Optimize performance on GPUs, 10x faster. Check them out: https://t.co/1CbPeRJhMY

[Aditya Rajagopal @adityaraja0]

@nCompass_tech is a performance optimization dev tool that unifies 𝗽𝗿𝗼𝗳𝗶𝗹𝗶𝗻𝗴, 𝘁𝗿𝗮𝗰𝗲 𝗰𝗼𝗹𝗹𝗮𝗯𝗼𝗿𝗮𝘁𝗶𝗼𝗻 and 𝘁𝗿𝗮𝗰𝗲 𝗮𝗻𝗮𝗹𝘆𝘀𝗶𝘀 of AI systems.

We automate the process of identifying and fixing the root cause of performance bottlenecks across all levels of the AI infrastructure stack. Our agent automatically surfaces key bottlenecks and works in tandem with Claude Code / Cursor to implement the fixes.

Using our tool, we were able to automatically improve a Hopper GEMM kernel taking it from 𝟯𝟬% 𝘀𝗹𝗼𝘄𝗲𝗿 𝘁𝗼 𝟯% 𝗳𝗮𝘀𝘁𝗲𝗿 than a near optimal CUTLASS kernel within a day - this took us months before!

We’re excited to officially release v0.1.0 of our VSCode extension that includes 𝗻𝗮𝘁𝗶𝘃𝗲 𝘁𝗿𝗮𝗰𝗲 𝘃𝗶𝗲𝘄𝗶𝗻𝗴, 𝘁𝗿𝗮𝗰𝗲 𝗱𝗶𝗳𝗳𝘀, 𝘁𝗿𝗮𝗰𝗲 𝗰𝗼𝗹𝗹𝗮𝗯𝗼𝗿𝗮𝘁𝗶𝗼𝗻 and integration of our 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁 directly into Cursor / Claude Code. We’re working closely with an initial set of users including the core maintainers of @vllm_project to build features that best support the community.

𝗜𝗳 𝘆𝗼𝘂’𝗿𝗲 𝗮𝗻 𝗲𝘅𝗽𝗲𝗿𝘁 𝗚𝗣𝗨 𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲 𝗼𝗽𝘁𝗶𝗺𝗶𝘇𝗮𝘁𝗶𝗼𝗻 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿 - bring your time spent optimizing performance down to days rather than weeks - view, collaborate and analyze faster than you do today.

𝗜𝗳 𝘆𝗼𝘂’𝗿𝗲 𝗻𝗲𝘄 𝗮𝗻𝗱 𝘄𝗮𝗻𝘁 𝘁𝗼 𝗹𝗲𝗮𝗿𝗻 𝗵𝗼𝘄 𝘁𝗼 𝗼𝗽𝘁𝗶𝗺𝗶𝘇𝗲 𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲 - you now have an expert pair-programmer to consult at any time.

𝗚𝗲𝘁 𝘀𝘁𝗮𝗿𝘁𝗲𝗱 𝗳𝗼𝗿 𝗳𝗿𝗲𝗲 𝘁𝗼𝗱𝗮𝘆 using VSCode, Cursor or Claude Code - https://t.co/lWXuYeoavi

We also work directly with you to identify opportunities for performance improvements. If you would like us to 𝗮𝘂𝗱𝗶𝘁 𝘆𝗼𝘂𝗿 𝘀𝘁𝗮𝗰𝗸 and suggest ways to improve performance - reach out at [email protected]!

[Aditya Rajagopal @adityaraja0]

Thanks! Yeah that's a great point. I agree with your point when you're optimizing arbitrary codebases, but I think given our current focus around AI model inference/training execution and the GPU kernels that get launched, the space for optimization is a bit more constrained - which makes our lives a little bit easier.

But this can still arise for instance when you say you want to optimize a kernel but you don't specify the input sizes you want to target. Something like a split-k optimization is only really valid if you're running a small input, so you wouldn't want to suggest that for large batch size kernels. We're building our agent to be aware of these trade-offs so that it asks for the constraints you are targetting before suggesting something that changes the structure of execution.

At that point it's up to you (if you're using this interactively) or the coding agent you're doing to specify the broader constraints better.

does that answer your question?

[Aditya Rajagopal @adityaraja0]

Diagnosing performance bottlenecks take 4–8x longer than fixing it.

We just shipped coding agent integrations for @nCompass_tech — now Claude Code and Cursor can talk directly to our agent to diagnose performance bottlenecks for you.

Add it as an MCP. Keep coding like normal.

Docs → https://t.co/GdYoTXqLo2

#CodingAgents #Vibetracing #CUDA #GPUKernels #ClaudeCode #Cursor

[Aditya Rajagopal @adityaraja0]

Anyone know if there's a maintained leaderboard for KernelBench outside of https://t.co/qI3Fsd5ZQY with more recent models and newer GPUs?

[Aditya Rajagopal @adityaraja0]

Last week we released AI agent integration for system level traces, i.e. nsys-reps and torch profile traces. Sharing a peak into what we're releasing soon for integrations with .ncu-reps for kernel developers!

Take a look here - https://t.co/mdQrmaHPMj for a demo of both and get started with agent assisted trace analysis.

For the first time you can run diffs and chat with all types of GPU performance related trace data all within your VSCode / Cursor IDE!

Excited to hear more around what parts of profiling and performance analysis you'd want to see more automation around. Reply in the comments below!

#MLSys #PerformanceOptimization #Profiling #nCompass

[Kai Lentit (e/xcel) @KaiLentit]

In 2026, AI models expire faster than session cache.

[Aditya Rajagopal @adityaraja0]

Got tired of manually reading through 200+MB trace files so we've now hooked them up to an AI agent and you can just chat with it. #vibetracing

Checkout our docs - https://t.co/GdYoTXqLo2

0:00 - Intro to the AI agent
1:00 - Example prompt with 2 questions
2:08 - Answer to question 1 - what should I care about in this trace
4:18 - Answer to question 2 - how do kernels map to operations from LLMs

Link to both the trace and responses in the comments!

#MLSys #GPUPerformance #PerfettoTraces #NsysReps

[Aditya Rajagopal @adityaraja0]

@3gilKarlsen @vllm_project @Alibaba_Qwen Thanks!

[Aditya Rajagopal @adityaraja0]

wrote a blog on how to use ncu directly on @vllm_project for a simple @Alibaba_Qwen model without taking 18+ hours to profile. I was tired of having to isolate kernels in a repro script to profile them with ncu. you can just isolate kernels direclty on a large trace and profile individual ones.

check it out here - https://t.co/fzJDABgTVA

#mlsys #vllm #inference #ai

[Aditya Rajagopal @adityaraja0]

@TheWake @vllm_project Nice! Yeah I agree, it's crazy we don't have a good way to view trace diffs in the GPU world as well.

[Aditya Rajagopal @adityaraja0]

we're introducing the concept of diffs to trace data - pick two traces, run the diff and get information on changes in durations, launch args, renames, additions and deletions of events in the trace.

check out how you can view diffs between @vllm_project 0.12.0 and 0.13.0 on the same workload - https://t.co/wjDtR8Dxhy

[Aditya Rajagopal @adityaraja0]

@Winterrose @grok who is this about?

[Aditya Rajagopal @adityaraja0]

Ever wanted to run Nsight Compute (ncu) on a large codebase and understand individual GPU kernel performance — without isolating each kernel?

I’ve hit this problem many times profiling systems like vLLM when asking the question - how well is a specific kernel performing for this workload?

In practice, that means using ncu. But with default settings, running ncu on thousands of kernel launches is prohibitively expensive — so you end up building isolated repos just to profile one kernel.

The good news: you don’t have to.

I’m writing a 3-part series on reducing ncu overhead enough to profile kernels directly in large, real-world codebases. Part 1 is out today.

Part 1 → understanding the metrics ncu reports (important because reducing which metrics you profile is important to reduce overheads of profiling)
Part 2 → how to actually apply this on large codebases
Part 3 → why combining ncu + nsys beats using either alone

📖 Part 1: https://t.co/lsoUXKWpWy

If you’ve run into this before, I’d love to hear in the comments how you handle kernel-level profiling at scale.

#CUDA #GPU #PerformanceEngineering #NsightCompute #MLSys #NVIDIA #Profiling

[Aditya Rajagopal @adityaraja0]

@marksaroufim @PrimeIntellect @modal @LambdaAPI Awesome! Yeah I've added the weekly meeting and I'm following the popcorn working group as well.