karthik.dev
Back to projects
Open Source

NVIDIA Agent Doctor

NVIDIA AI environment diagnostics, security, compatibility & benchmarking CLI with MCP integration

GitHub
PythonNVIDIA CUDAPyTorchGPU DiagnosticsMCP ServerDeep LearningCLI

Overview

An open-source CLI that automates GPU, CUDA, and PyTorch compatibility checks, configuration audits, security scans, and bounded benchmarking across NVIDIA AI environments. Designed to make local AI-environment troubleshooting reproducible, auditable, and integrable with MCP server tooling.

Problem

Setting up NVIDIA GPU environments for AI workloads involves manual compatibility checks across driver versions, CUDA versions, PyTorch builds, and security configurations. Diagnosing failures is time-consuming and non-reproducible — engineers repeat the same debugging steps for every environment.

Solution

A structured diagnostic CLI that runs reproducible checks across the full GPU software stack: driver compatibility, CUDA version alignment, PyTorch installation validation, security configuration audits, and bounded performance benchmarks. Outputs structured reports for auditability.

Architecture

  1. 1GPU hardware detection and driver version extraction
  2. 2CUDA version compatibility matrix validation
  3. 3PyTorch installation and CUDA binding verification
  4. 4Security configuration audit — exposed ports, unsafe configurations
  5. 5Bounded benchmarking — deterministic performance measurement without runaway GPU utilization
  6. 6Structured report output — JSON and human-readable formats
  7. 7MCP server integration for agent-driven environment diagnostics

Technical Challenges

  • Designing bounded benchmarks that measure GPU performance reproducibly without saturating resources or causing instability.
  • Building a CUDA compatibility matrix that remains accurate as NVIDIA driver and CUDA versions evolve.
  • Structuring diagnostic workflows that surface actionable fixes, not just raw version numbers.

Results

  • Open-sourced NVIDIA AI environment diagnostics tool — instantly usable by any AI engineer.
  • MCP server integration enables agent-driven AI environment troubleshooting.
  • v1.0.0-alpha released with full GPU, CUDA, PyTorch, security, and benchmarking modules.