CaptionDoor

Inspiration

A door visit should not depend on hearing a bell or a spoken message. CaptionDoor explores a caption-first interface with clear responsibility, readable text and an optional visitor page.

What it does

A resident sees a door alert, explicitly claims responsibility, reads revisable captions and can hand the visit to another resident only after acceptance. Visitors can separately choose a browser text conversation without an account. Replies distinguish ACCEPTED, DISPLAYED and explicitly CONFIRMED states. Presets draft a reply; they do not send it. Closing a visit records an outcome rather than claiming that someone has left.

How we built it

TypeScript, Next.js, React and a dedicated WebSocket/job worker. DynamoDB is the only persistent application database, with conditional writes, tenant-scoped keys, TTL, indexed due work and explicit content deletion. Browser-safe contracts are separate from database and integration SDKs. The reproducible local demo uses real DynamoDB Local and clearly labeled simulated events and synthetic captions.

Ring and AWS integration: observed status

The repository implements Ring account/device discovery, configuration and receive-only WHEP integration with provider consent and bounded session controls. An authorized temporary Ring Playground token successfully discovered an authorized device in a separate bounded test. That test did not observe incoming media, and AWS Transcribe was not started. The video is explicitly a LOCAL/FIXTURE rehearsal, not a recording of live Ring audio or AWS-generated captions. No physical Ring device success is claimed.

A separate AWS DynamoDB development stack with three dedicated tables was created and its metadata, TTL and bounded fictional writes were verified. Exact owned visit/runtime data and encrypted-token-copy absence were verified after the test; the owned services stopped. Ring DELETE/provider expiry was not independently observed. AWS Transcribe integration is implemented and gated but end-to-end transcription remains unverified.

Demo and testing

The English on-screen, silent demo shows the simulated bell, claim, Updating/Finalized synthetic captions, media expiry, accepted handoff, optional visitor text, explicit Send, delivery acknowledgement and closure. Judges can reproduce the local workflow without supplying a Ring token: follow README local setup, start only the captiondoor_local Compose project, then run corepack pnpm judge:replay. Default demo sources remain clearly labeled. Real-provider testing is separate and opt-in.

AI assistance

OpenAI Codex assisted implementation, debugging, test development and documentation. The local video uses deterministic fixture captions; it does not depict an AI transcription model operating on real incoming audio.

Challenges and learning

We separated responsibility from connectivity, recognition finality from truth, delivery from acknowledgement and fixture evidence from provider evidence. Real-provider verification exposed the importance of preserving sanitized HTTP diagnostics rather than treating every missing-track observation as the same failure.

Next steps

Resolve and re-test Ring media negotiation under a separately approved bounded attempt, then verify actual incoming-audio-to-AWS-final transcription and real device/browser compatibility.

Built With

• amazon-dynamodb
• aws-sdk
• docker
• next.js
• openai-codex
• playwright
• react
• ring-api
• typescript
• websocket