Surplus Switchboard

Surplus Switchboard helps a kitchen operator explore capacity inquiries and constrained surplus-food allocation.

Working implementation

The local Python application computes exact integer allocation subject to compatibility, expiry, receiving capacity and operator clearance. A guarded CALL-E REST workflow binds a specifically authorized request to a digest, records the attempt before creation, and blocks automatic redial after uncertainty. Review and reconciliation code binds evidence to its original planning snapshot. The browser itself has no live-call endpoint.

Actual CALL-E evidence

One explicitly authorized recorded AI role-play completed on September 13, 2026. The authenticated provider dashboard shows 46 seconds, ended by the agent. The recipient acknowledged AI recording, accepted six fictional produce portions, and confirmed the read-back. This demonstrates an actual telephone interaction using fictional data, not a real organization or donation.

The create request did not return a confirmed logical API task ID. The dashboard recording ID returns 404 from the documented REST lookup. The durable ledger remains locked and no duplicate call was placed. Automated provider-result-to-planner reconciliation is therefore incomplete. The entry intentionally discloses this limitation; no fabricated structured result is used.

Request and support correspondence

The repository's submission folder includes the sanitized actual API request, real transcript excerpt, and sanitized outgoing email sent to CALL-E support requesting read-only task recovery. The sent email was verified in the entrant's mailbox. No support reply or endorsement is claimed. Private phone numbers, credentials, signed audio URLs and handoffs are excluded.

https://github.com/Bembaby/surplus-switchboard/tree/main/submission

Verification

133 tests pass from the publication checkout. They cover allocation, call creation safeguards, bounded result reads, evidence binding, role-play scope and HTTP behavior. Provider fixtures are not live evidence. The deliberately constructed local example proposes twelve portions versus six for nearest-first; this is not measured food-waste reduction or a novel algorithm.

The real call exposed a disclosure issue after phone screening: the agent used the operator's name after an interrupted AI introduction. The recipient nevertheless expressly acknowledged AI status. A local prompt correction now requires repeating AI/recording/fictional disclosure and forbids identifying as the operator. Regression tests pass, but no second live call validates the correction.

Demo and contribution

The public video contains continuous pre-call local planner footage, followed by a card explaining the subsequently verified real call and unresolved REST recovery. It is English-captioned, without voice narration or live-call audio. It does not portray a completed automated integration.

Contribution: https://github.com/CALLE-AI/awesome-phone-call-agents/pull/368

Boundaries

No real food is reserved, delivered or certified safe. This is a local single-operator prototype. CALL-E support recovery and a verified end-to-end result import remain future work. Codex assisted implementation, tests, documentation and submission preparation. Belal Embaby is the solo entrant; source is MIT licensed.