Overview of POC Factory#

POC Factory turns requirements, proposals, and RFPs into downloadable AI and Physical AI proofs of concept with code, tests, and deployment instructions. Use the browser application to review a design, generate an implementation, and inspect the validation evidence before running the result.

Use this documentation to understand the application, install it from the public NVIDIA NGC catalog, and generate your first proof of concept. For changes in the latest release, see Release notes and the Upgrade guide.

Why POC Factory Exists#

A useful proof of concept needs more than generated source files: it needs a clear scope, an architecture, tests, packaging, and evidence of which checks ran. POC Factory brings those steps into one workflow built on NVIDIA NeMo Agent Toolkit, with optional human review before implementation.

Blueprint discovery can match requirements to 42 curated NVIDIA blueprints. An OpenAI-compatible provider supplies the chat and embedding models; models are external services rather than weights bundled in these containers. You do not need a local GPU to run the POC Factory application stack.

Common Use Cases#

Use the application for these common workflows. Each may need additional credentials or infrastructure beyond the base installation.

Goal	Starting point	What to expect
Build an AI agent POC from a proposal	Create → AI Agent → Proposal to POC	Requirements, design artifacts, implementation, tests, and a downloadable package
Review a design before generation	Enable Human-in-the-Loop Mode	Review pauses for the PRD, blueprint selection, and architecture
Explore Physical AI	Create → Physical AI	Physical AI workflows and OpenUSD assistance; downstream simulation has its own requirements
Work with a completed POC	Accelerate	Deployment, evaluation, profiling, optimization, monitoring, and sizing actions
Prepare model customization	Create → model fine-tuning	Dataset and customization workflows against configured NeMo Microservices
Draft a proposal from an RFP	RFP	Proposal content that can inform a later POC workflow

Core Concepts#

Understand these terms before choosing the deployment and generation settings.

Concept	Meaning
Release set	Backend, frontend, and GitHub MCP proxy images selected together from the same release, using exact tags or approved digests
Inference profile	Provider base URL, API key, and five model roles: code, reasoning, balanced, fast, and embedding
Contract-first generation	The default `auto` path authors a capability contract, builds the application scaffold and tests, and uses Cursor for bounded implementation edits
Validation evidence	Recorded checks and failures, including explicit skipped gates; a skipped Docker or runtime gate is not a pass
Durable state	Database records, generated packages, job state, saved credentials, and optional fine-tuning data retained across restarts

Core Components#

Deploy the application as a coordinated stack. The detailed connections and state boundaries are described in Architecture.

Component	Purpose
Frontend	React UI served by nginx; proxies API and streaming progress to the backend
Backend	FastAPI APIs and NeMo Agent Toolkit generation, validation, and packaging workflows
GitHub MCP proxy	Optional authenticated source and blueprint intelligence; included in the standard Compose profile
PostgreSQL	Persistent users, POC records, and encrypted settings
Phoenix	Optional tracing and observability