NVIDIA Corporation

Investigate failures in NVIDIA TAO Visual ChangeNet classification experiments using prediction scores and actual images. Produces an illustrated root cause report covering data, reference images, training settings, and simulated changes to assess possible fixes.

• Investigate missed defects and poor recall
• Analyze FAR regressions and false positives
• Audit reference images and inspection data

Replace the object detector in VSS Alerts Blueprint verification mode with an ONNX model. This guide covers model placement, interpreting detection outputs, DeepStream configuration, TensorRT engine builds, redeployment, and common loading or mount failures.

• Replace the verification detector with a YOLO11 ONNX model
• Investigate DeepStream custom parser loading failures
• Fix ONNX mounts and TensorRT engine rebuild settings

Fine-tune the MAISI VAE, a model that compresses and reconstructs CT or MRI volumes, using NIfTI datasets. Prepare training settings, check inputs before training, and save model checkpoints and loss histories for review.

• Check a NIfTI dataset before GPU training
• Fine-tune a VAE with CT or MRI volumes
• Compare reconstruction losses across training runs

Convert existing PyTorch training code into an NVFLARE federated learning job, where sites train with their own data. Adapt training and evaluation, exchange model updates through the Client API, validate locally, and export the job.

• Convert a manual PyTorch training loop to federated learning
• Adapt existing evaluation code for federated rounds
• Validate and export a FedAvg training job

Generate videos from DEFT media samples using PAIDF Cosmos Predict. Prepare configuration, run generation in Docker, avoid repeated processing of duplicate media paths, and produce JSONL records linking each original sample ID to successful or missing video outputs.

• Generating synthetic videos for DEFT media samples
• Avoiding repeated generation for duplicate media paths
• Passing generated video mappings to downstream workflows

Diagnose slow loading, high memory use, and low frame rates in USD 3D scenes. It coordinates measurement, scene validation, and optimization, then produces reports comparing available before-and-after evidence and an optimized scene when changes are performed.

• Investigate slow loading in USD scenes
• Diagnose low frame rates and high memory use
• Review repeated structures in large CAD scenes

Evaluate object detection predictions against labeled ground truth with TAO Data Services. Produce class-level precision, recall, accuracy, and AP, a detection quality measure, with CSV results and precision–recall plots for reviewing model performance.

• Score object detection predictions against ground truth
• Review precision and recall by class
• Generate metrics CSV files and precision–recall plots

Find the right next step for an unclear NVFLARE request. Uses source or data inspection evidence to recommend a specific workflow skill for conversion, validation, deployment, or diagnosis, while keeping the project unchanged.

• Choosing where to start with NVFLARE
• Mapping a local project to a FLARE workflow
• Resolving unclear or conflicting training ownership

Coordinate defect image generation, augmentation, and labeling for circuit board, metal surface, and glass inspection on OSMO. Select workflows that create initial board datasets from CAD scenes, process real images, or fine-tune NVIDIA Cosmos AnomalyGen models.

• Creating initial circuit board inspection datasets from CAD scenes
• Generating circuit board texture and component pose defects
• Running inference and labeling on real inspection images

Turn records from a stopped NVFLARE Auto-FL optimization campaign into a reproducible Markdown report and JSON summary. Summarize improvements, retained candidates, failed ideas, and literature outcomes, with a refreshed progress plot when plotting dependencies are available.

• Create a final report after an optimization campaign stops
• Summarize baseline scores and retained improvements
• Review successful and unsuccessful literature-based ideas

Find PhysicsNeMo models, data loading pipelines, and examples for machine learning tasks involving physical systems. The skill checks current repository files, presents relevant alternatives, and suggests a reading order without writing training code.

• Finding model families for physical forecasting
• Comparing models for grids and unstructured meshes
• Finding data loading pipelines for a dataset

Convert existing Hugging Face Trainer or TRL SFTTrainer code into an NVFLARE job for federated learning, where multiple sites train a shared model. Preserve training settings, validate locally, and export the resulting job.

• Converting Trainer-based training into a federated job
• Adapting PEFT or LoRA fine-tuning for federated learning
• Preserving evaluation metrics during conversion

Evaluate answer quality in NVIDIA RAG Blueprint using local documents and question data. The skill guides dataset preparation, RAGAS scoring, configuration comparisons, and error diagnosis for RAG systems that retrieve documents to support generated answers.

• Prepare documents and questions for evaluation
• Compare quality across retrieval and generation settings
• Interpret RAGAS result files

Fine-tune NV-Generate-CTMR MR-Brain v1 using supplied brain MRI volumes. Check the input file list, prepare training configurations, run training, and optionally generate sample volumes, producing a saved model and a structured run summary.

• Check a brain MRI dataset before GPU training
• Fine-tune a model with T1, T2, FLAIR, SWI, or MRA data
• Prepare a training command with explicit data and output paths

Optimize stock allocations from price data using NVIDIA cuOpt, with controls for severe losses or overall variability. Compare risk and return, inspect allocation tables, backtest against benchmarks, and evaluate scheduled or drift-triggered rebalancing.

• Calculate allocations under risk and asset weight limits
• Compare allocations across risk preferences
• Backtest an allocation against an equal-weight portfolio

Run NV-Reason-CT on a supplied 3D chest or abdominal CT scan and a text prompt for research or engineering. Returns generated text and JSON records of input geometry, model identity, dependencies, and limitations.

• Testing inference on a chest CT scan
• Running abdominal CT inference with a focused question
• Checking local model assets and runtime readiness

Fine-tune NV-Segment-CT VISTA3D on CT images and region labels in NIfTI format. Choose standard training or fixed-class softmax training, save model checkpoints and evaluation records, and optionally track training with MLflow.

• Fine-tuning with paired CT images and region labels
• Checking the training workflow with a smoke run
• Training mutually exclusive classes with softmax

Convert existing PyTorch Lightning training code into an NVFLARE job for federated learning, where sites train collaboratively while keeping data local. It integrates model exchange through the Lightning Client API and supports local validation or job export.

• Converting Lightning training code to federated learning
• Preparing collaborative training with data kept at each site
• Checking a converted job in a local simulation

Set up NVIDIA infrastructure for synthetic data generation on MicroK8s or Azure AKS. Coordinate Kubernetes, inference services, and OSMO workflow execution, with prerequisite checks, deployment validation, capacity planning, and failure investigation before workloads are retried.

• Set up a synthetic data generation environment on MicroK8s
• Deploy an Azure AKS environment with OSMO and inference services
• Check GPU quotas and capacity before scaling

Provides shared guidance and templates used by NVFLARE conversion skills. Supports consistent dependency setup, site-specific data paths, validation, and custom model aggregation, combining results from participating sites. It is an internal companion rather than a standalone converter.

• Consulting dependency setup rules during conversion
• Checking data paths for participating sites
• Following shared validation and reporting guidance

Find the right NVIDIA NuRec skills for turning sensor recordings into renderable 3D scenes. This routing guide organizes dataset downloads, conversion, training, rendering, and frame cleanup into workflows handled by separate specialist skills.

• Choose the right NuRec skill
• Plan 3D reconstruction from sensor recordings
• Render a pretrained NuRec scene

Deploy and manage NVIDIA RAG Blueprint, a system that retrieves information to support AI answers. Check the running environment, configure models and search features, troubleshoot unhealthy services, and verify affected workflows after changes.

• Deploy NVIDIA RAG Blueprint
• Configure search and document ingestion
• Change models or enable image features

Segment specified organs in CT NIfTI scans using NV-Segment-CT VISTA3D. Produces a label map identifying organ regions and JSON evidence covering labels, voxel counts, physical volumes, geometry checks, and model identity for engineering validation.

• Segmenting the liver, spleen, and both kidneys in CT scans
• Recording voxel counts and physical volumes for each organ
• Checking output geometry and unexpected labels

Create or refresh a governance card, a structured review document for an existing agent skill. It gathers local source evidence, renders a Markdown draft, and checks for unresolved review markers before submission for human review.

• Create a governance card for an existing skill
• Refresh a card after skill changes
• Prepare material for legal and safety review