Use POC Factory#

Users work through the deployed web application. Start an AI agent POC from requirements or a proposal, review optional Human-in-the-Loop decisions, then inspect the packaged result and continue in Accelerate.

Prerequisites#

Before starting, obtain the application URL and sign-in access. Check Settings for an effective inference profile. The default auto generation path also needs a validated Cursor credential; an operator can instead select the separate llm generation mode.

Navigate the Application#

Use the top-level areas for different stages of work.

Area	Purpose
Dashboard	View POC inventory, job status, and resumable work
Create	Select an AI Agent, Physical AI, or model fine-tuning workflow
Generate	Submit AI agent requirements and optional review gates
Accelerate	Deploy, evaluate, profile, optimize, monitor, or size a completed POC
Finetune	Prepare datasets and manage customization workflows against configured NeMo Microservices
RFP	Generate proposal content from an RFP
Settings	Review account and permitted inference or Cursor credentials

Start and Review a Generation#

Follow these steps for an AI agent POC.

Select Create, choose AI Agent, then choose Proposal to POC.
Enter implementation-ready requirements. Describe users, outcomes, inputs, outputs, integrations, failure behavior, privacy constraints, and validation checks.
Optionally name the project and choose a domain. Leave domain auto-detection on for an initial run.
Enable Human-in-the-Loop Mode when the PRD, blueprint choice, or architecture needs explicit review.
Select Generate POC and monitor the Progress page or its Dashboard entry.
At a review pause, approve or request a revision, then continue the same job from Dashboard. Avoid submitting a duplicate request for the same paused job.
When packaging completes, review validation results and download the archive.

For a concrete first request, follow Quickstart for POC Factory.

Understand the Generation Stages#

The nine-stage workflow produces these artifacts or checks. Blueprint selection may be skipped when no blueprint is used.

Stage	Result	User action
Requirements	Interpreted scope and constraints	Review in Human-in-the-Loop mode
PRD	Product requirements document	Approve or request revision when prompted
Blueprint	Search, ranking, and selection when useful	Confirm or choose when prompted
Architecture	Components and interfaces	Approve or request revision when prompted
Code	Application implementation and packaging inputs	Monitor
Tests	Generated or manifest-derived tests	Monitor
Validation	Configured checks and explicit skipped gates	Review failures and skips
Documentation	README and operating notes for the generated POC	Inspect with the package
Packaging	Downloadable archive and durable project record	Download and inspect

The deployment’s validation policy determines whether Docker build and runtime checks run. A skipped gate is recorded as skipped, not passed. See Configure and validate.

Work with a Completed POC#

Open a completed POC in Accelerate for lifecycle actions.

Tab	Use it for
Deploy	Test SSH access and deploy a generated Docker workload
Evaluate	Measure response or retrieval quality
Profile	Inspect latency, token usage, spans, and bottlenecks
Optimize	Tune workflow configuration
Monitor	Review health and telemetry
Scale	Estimate resource needs from workload observations

Physical AI guidance and OpenUSD assistance start through Create. For remote model customization, follow Fine-tune a model; the deployment operator can connect existing NeMo Microservices endpoints.

Next Steps#

Read Operate and restore for job and artifact retention, or Troubleshooting if progress, provider validation, or download fails.

For an administrator Cursor key, declare MODEL_CURSOR=auto or a supported model identifier. Read Architecture for the generation and state boundaries, and Upgrade guide when updating an existing deployment.