nCompass Technologies
nCompass Technologies Inc. · Writing · Voice & Language · Other
nCompass Technologies makes an AI agent and VS Code extension that read GPU traces, spot CPU-GPU bottlenecks, and suggest fixes for AI inference stacks.
About nCompass Technologies
What Is nCompass Technologies
nCompass Technologies is a Y Combinator backed startup behind a performance profiling tool built for the people who keep AI systems fast. Its current product pairs an AI agent with a lightweight VS Code and Cursor extension, so you can open a trace, ask why a serving job stalls, and get an answer that points at a specific kernel, a memory copy, or a scheduling gap.
The company started in 2024 as an inference hosting platform with a rate-unlimited API. That business pushed the team into the same problem their customers kept hitting: the serving stack was fine on paper and slow in production, and nobody could say where the time went. The tool they built to answer that question turned into the product they now sell, aimed at AI platform teams running generative recommender, ranking, and LLM serving at scale.
The premise is simple to state and hard to do. Finding the cause of a performance problem takes four to eight times longer than writing the code that fixes it. Engineers copy a several gigabyte trace to a laptop, scroll through a wall of counters, and guess. nCompass compresses that loop into an editor window and a conversation.
The limits matter as much as the pitch. This is a tool for teams that already run NVIDIA profilers and can reproduce a slow workload on demand. If you have no trace, there is nothing to analyze. Small projects with one model and no latency target will not get much from it, and the paid side of the product is arranged with sales rather than a checkout page.
Getting Started
- Create an account at ncompass.tech and install the extension from the VS Code marketplace, or add it to Cursor.
- Profile your workload with Nsight Systems or the PyTorch profiler so you have a trace file to work with.
- Open the trace in the extension, or push it to the cloud with the npx ncompass command so teammates can open it too.
- Ask the agent a direct question about the run, such as where the idle gaps between kernels come from, and read the ranked list of suspects.
- Compare the new run against the old one with TraceDiff to confirm the change actually moved the numbers.
Product Information
A quick look at nCompass Technologies's pricing, supported platforms, and performance.
Best for
The users, tasks, and scenarios where this tool fits best.
Users
- ML platform engineers
- GPU and inference engineers
- Engineering leads scoping performance work
- Kernel developers writing CUDA
Tasks
- Finding why a serving job idles
- Comparing two runs after a config change
- Reviewing a trace with a teammate
- Prepping a kernel rewrite
Scenarios
- A latency regression shows up after a framework upgrade and the team has a week to find it
- A generative recommender stack has no public tuning community
- You are deciding whether a GPU fleet needs more cards or better code
Key features
AI Agent That Reads Your Traces
The agent takes a profiling file plus the source code around it and returns a short list of bottlenecks with reasoning attached. It works at two levels: system wide questions about scheduling and data movement, and kernel level questions about occupancy, memory access, and instruction mix. You can push back on its answer and ask it to check a different hypothesis, which is closer to working with a colleague than running a report.
Trace Viewer Inside the Editor
Nsight Systems and PyTorch profiler output both open inline, so the trace never has to leave the machine you are working on. You get the timeline, the kernel list, and the source pane in one window, and jumping from a slow kernel to the line that launched it takes a click. For anyone who has spent an afternoon scrolling a browser based trace viewer, the difference is mostly about not losing your place.
TraceDiff Run Comparison
TraceDiff puts two runs side by side and highlights what changed: which kernels appeared, which ones got slower, and where the gaps grew. This is the feature that answers the question teams ask most often after a deploy, and it answers it with data rather than a hunch.
Sharing and Collaboration
Traces can be pushed to the cloud with the npx ncompass command, listed, pulled, and shared through a link. A colleague opens the same view you see, with the same zoom level and the same annotations. That sounds minor until you try to debug a performance problem over a chat thread.
Kernel Level Analysis
Kernel work is where most of the time goes in a slow AI system, and it is also where general purpose assistants fall apart. The agent reads ncu style counter data, detects repeating kernel patterns across a run, and proposes concrete changes such as tiling, vectorized loads, or a different launch configuration. The company says its own team used the tool to build a Hopper GEMM kernel that beat NVIDIA CUTLASS by about 3 percent in a day, a job that took months before.
Works Inside Cursor and Claude Code
The agent connects to the editors engineers already use, so the reasoning and the code edit happen in the same place. You can ask for an analysis in Cursor, get the diagnosis, and have the patch drafted without switching tools. For teams standardizing on AI assisted development, that removes the usual gap between a profiling report and the commit that acts on it.
Pros and cons
Pros
- Collapses trace reading, sharing, and diagnosis into one editor window, which cuts the slowest part of a performance investigation.
- Handles both system level and kernel level questions, so you do not need a second tool for CUDA work.
- TraceDiff gives a defensible answer to "did this change help" without re-running a full benchmark suite.
- Free to start with the VS Code extension, which makes it cheap to test on a real workload before talking to sales.
Cons
- Useless without a trace. If your workload cannot be reproduced or profiled, the agent has nothing to reason about.
- Built around NVIDIA tooling, so teams on other accelerators will find the workflow does not line up with their stack.
- The paid tiers and team features are sales led, which means no published price and a slower path to a purchase decision.
Frequently asked questions
An AI agent and editor extension for GPU performance work. It reads profiling traces from AI serving and training workloads, points at the bottlenecks, and suggests code changes. The company also offers engineering help on a project basis.
Related content
Explore related tools, skills, and articles for nCompass Technologies.
nCompass Technologies Alternatives
Google EmbeddingGemma 2
Google DeepMind · Writing · Image · CodingGoogle EmbeddingGemma 2 is a 740M-parameter open model that turns text, code, images, audio, and video frames into vectors in one 768-dimension space, running on phones and laptops.
MePersona
MePersona, Inc. · Writing · Productivity · MarketingMePersona is an AI unified inbox for creators. It learns your voice, drafts replies to fan messages from 12 platforms, and sends only what you approve.
Fluence
Yay Software (Yannick Huchard) · Writing · Voice & Language · ProductivityFluence is an Android app that reads PDFs, EPUBs, Word files and web articles aloud, follows each sentence as it speaks, and answers questions about the text on the phone itself.