AI-Assisted Development

This page demonstrates AI-assisted development: describe the requirement in one sentence of natural language, and Claude (using the ne503-dev skill) takes care of the app's development, deployment, and verification — with no manual command execution.

The demo app is a loitering alert (Loitering Detection): when a person stays in frame continuously for 10 seconds, it fires an alert; when they leave, it resets.

Try the finished app directly

Want to skip development and deploy the finished app right away? Download the prebuilt package lingering-detection.tar, unzip it to get app.yaml and image.tar, and follow the steps in Hello World §4 to deploy to the device.

1. Prerequisites

Prerequisite	Description
Claude Code	Install it on your local machine.
ne503 source repo	Clone locally (bundled with the NE503 SDK); the ne503-dev skill resides at .claude/skills/ne503-dev/. Alternatively, download the skill zip and extract it into .claude/skills/.
Docker	Install it on your local machine, for building app images.
NE503 device	Prepare a ready device: note its IP + admin password, and confirm the platform is initialized (HALv2 installed, ai-runtime healthy, detection model scanned + loaded).

2. From One Sentence to a New App

Traditional development requires writing the app.py business logic, configuring app.yaml permissions, discovering the device's available models and video streams, building the image, deploying, and verifying — AI-assisted development hands all of this to Claude; the developer only describes the requirement in natural language. Below is a real on-device session (2026-06-22) where the only input was a single sentence.

2.1 Input requirement

In Claude Code, invoke the ne503-dev skill and describe the requirement in natural language:

Build an app: fire an alert after someone is detected lingering for 10 seconds. Deploy to <device-IP>.

Claude reads the skill, starts planning on its own, and — after confirming a few details with the developer — gets to work:

2.2 Claude develops the app (the core)

Once the requirement is confirmed, Claude develops the app autonomously, starting from the SDK's apps/template/ app template. Three key decisions shape the app:

• Placeholder fixes: the template's sample model and stream don't run on a real device; Claude queries the device and substitutes the real values — model hailo_yolov8n_384_640, stream sub (publishes raw NV12 frames; main only publishes encoded H264 and cannot be used for inference).
• Dwell state machine: the app's core logic. A timer starts when a person enters the frame; an alert fires once they linger continuously for 10 seconds; if no person is detected for 3 seconds, they're deemed to have left and the state resets. A 3-second grace window (GRACE_SECONDS) tolerates brief detection drops from turning or occlusion.
• Manifest config: app.yaml declares the required video stream, model, event topics, and tunable env vars (dwell seconds LOITER_SECONDS, detection threshold DETECTION_THRESHOLD, grace seconds GRACE_SECONDS, etc.).

The full development process (template selection → placeholder fixing → state machine implementation) is below:

2.3 Claude deploys and verifies

With the code written, deployment is likewise handled by Claude — the skill's bundled script chains "build → upload → install → start → verify" into one fully-automatic pass, and the app enters running.

After deployment, Claude verifies three things: inference is actually live (logs confirm the model is loaded and the first frame has arrived); the platform-injected permissions match app.yaml; and a person entering the frame triggers a complete detection cycle. The verification from the Web Console's perspective is below:

The cycle repeated many times on-device; every alert landed at 10.1 s, totaling 3 independent alerts — no false alarms, no missed ones.

A single sentence of input produces a live loitering-alert app on the device — with no manual step in between.

Full end-to-end process (watch on demand):