NVIDIA Digital Health Examples

Agent skills and worked examples for healthcare AI workflows from NVIDIA Digital Health.

This repository hosts agent skills — markdown-based workflow guides consumable by AI coding assistants (Claude Code, Cursor, GitHub Copilot, and any client that supports the agentskills.io specification).

The first set covers a clinical ASR (automatic speech recognition) evaluation workflow: term curation, synthetic clinical-speech benchmark generation, KER (Keyword Error Rate) / entity-level scoring, and fine-tune guidance.

These skills are docs-only workflow guides. They do not ship companion runtime code; an agent following them will use the user’s own environment and any required NVIDIA APIs (NIMs at build.nvidia.com, NeMo containers, etc.).

Skills in this repo

Each skill folder contains SKILL.md (the workflow guide), optional references/*.md (deeper detail loaded on demand), and evals/evals.json (trigger / behavior / boundary test cases).

Getting started

Skills are loaded by your agent’s skill loader. For Claude Code:

git clone [email protected]:NVIDIA/digital-health-skills.git
cd digital-health-skills

# Either symlink or copy the four skill folders into your Claude Code skills dir:
mkdir -p ~/.claude/skills
ln -s "$(pwd)/skills/digital-health-clinical-asr-setup"    ~/.claude/skills/
ln -s "$(pwd)/skills/digital-health-clinical-asr-build"    ~/.claude/skills/
ln -s "$(pwd)/skills/digital-health-clinical-asr-eval"     ~/.claude/skills/
ln -s "$(pwd)/skills/digital-health-clinical-asr-finetune" ~/.claude/skills/

Then invoke them from your agent: /digital-health-clinical-asr-setup, /digital-health-clinical-asr-build, /digital-health-clinical-asr-eval, /digital-health-clinical-asr-finetune.

The skills walk the agent through each stage and hand off to the next. Start with /digital-health-clinical-asr-setup.

Requirements

To use the skills end-to-end you will need:

• An NVIDIA_API_KEY from build.nvidia.com — for hosted Magpie TTS + Parakeet/Nemotron ASR via NVCF (free tier is sufficient).
• An NGC_API_KEY from ngc.nvidia.com — only needed for Stage 4 (fine-tune), to pull the NeMo training container.
• Python 3.10+ for client-side scripting (the agent installs deps via your skill of choice).
• A CUDA host with the NeMo container (nvcr.io/nvidia/nemo:25.11.01) for Stage 4. Brev cloud GPUs supported.

Support

• Level: Experimental — the skills are new and evolving; expect rough edges.
• Issues: Use this repository’s Issues for bug reports and feature requests. Do not file security issues there (see SECURITY.md).

Contributing

This is a reference samples repo maintained by the NVIDIA Digital Health team — external pull requests are not accepted. See CONTRIBUTING.md for the full policy and what to do if you found a bug (open an Issue) or want to adapt the skills for your own use (fork under the Apache 2.0 license).

Security

See SECURITY.md for the NVIDIA vulnerability disclosure process. Report security issues to [email protected], not via GitHub.

License

Apache License 2.0 — see LICENSE.