Building domain-specific AI agents for hardware reliability at NVIDIA
In this GitHub Universe 2026 session, Maria Victoria Gromysch, Test Engineer at NVIDIA, presents a practical framework for applying AI throughout the engineering lifecycle, with a focus on hardware yield and reliability.
AI across the engineering lifecycle
The session describes how NVIDIA combines GitHub Copilot, GitHub Actions, and custom AI agents at different stages of development:
- Pre-commit: AI-assisted test generation.
- Commit: AI-assisted code reviews.
- CI/CD: Automated validation.
The central distinction is between catching surface-level code issues and understanding hardware-specific constraints, firmware interactions, and system-level failures. General-purpose models can miss these domain-specific concerns, so the session covers training agents on internal diagnostics to support hardware reliability work.
Agent design and domain knowledge
Gromysch covers the motivation behind NVIDIA’s AI agent, its candidate selection and operations, how it works, and its impact and benefits. The discussion also addresses data ingestion and processing, project challenges, reinforcement learning (RL) systems, machine learning approaches, and data storage.
Domain knowledge is a key theme: the session examines what makes the agent specific to NVIDIA’s hardware reliability needs and how internal diagnostics help address the limits of general-purpose AI models.
Session chapters
- 00:00 — Introduction to NVIDIA’s AI agent
- 00:18 — Understanding hardware yield and motivation
- 01:13 — Agent design: candidate selection and operations
- 02:34 — How the NVIDIA AI agent works
- 04:41 — Impact and benefits of the AI agent
- 06:26 — Q&A: data ingestion and processing
- 08:32 — Q&A: project challenges and RL systems
- 11:46 — Q&A: uniqueness and domain knowledge
- 13:28 — Q&A: ML approaches and data storage