opendoor call-e

Inspiration Accessibility information for local businesses is often missing, outdated, or hidden inside reviews. For wheelchair users, caregivers, and families, this uncertainty can make a simple visit stressful. We created OpenDoor to turn phone conversations into trustworthy, reviewable accessibility information. What it does OpenDoor uses CALL‑E to contact authorized businesses and verify their current opening hours and wheelchair accessibility. The agent identifies itself, requests consent, asks only approved questions, and captures exact supporting evidence. Results are never published automatically. Unsupported or low-confidence answers are quarantined, while valid updates require human review and approval. How we built it We built OpenDoor with TypeScript, React, Vinext, Cloudflare Workers, D1, and the CALL‑E API. The workflow:

1. Checks authorization, suppression rules, quiet hours, and destination details.
2. Creates an idempotent call to prevent duplicate dialing.
3. Requests consent before collecting information.
4. Retrieves CALL‑E’s structured result and transcript evidence.
5. Rejects conflicting, incomplete, or unsupported claims.
6. Presents valid changes for human review.
7. Publishes only explicitly approved updates.
8. Records an audit trail showing what changed and why. We also built deterministic simulations, integration tests, browser tests, evaluation scenarios, and secret-scanning safeguards. Challenges we ran into The hardest challenge was making an AI phone agent trustworthy, not merely conversational. Network failures can leave call status uncertain, retries can accidentally create duplicate calls, and generated summaries may contain claims that are not supported by the conversation. We addressed these problems with durable call intents, idempotency controls, fail-closed validation, exact evidence matching, confidence thresholds, and human approval. Accomplishments that we're proud of We are proud that OpenDoor treats safety as part of the product rather than an afterthought. It can prevent duplicate calls, respect denied consent, preserve existing information when verification fails, and clearly separate verified evidence from uncertain claims. We also created a complete review experience where operators can inspect evidence and approve individual facts before publication. What we learned We learned that responsible voice automation requires more than a strong prompt. It needs consent, evidence, authorization, deterministic safeguards, recovery paths, and human accountability. We also learned how CALL‑E can connect offline information with digital services while keeping people in control of the final decision.

Built With

• ai-voice-agents
• base-ui
• call-e
• cloudflare-d1
• cloudflare-workers
• drizzle-orm
• human-in-the-loop
• node.js
• playwright
• react
• shadcn/ui
• tailwind-css
• typescript
• vinext
• vite
• wrangler