Future Call AI — Evidence-First Autonomous Calling

An evidence-first AI phone agent that turns CALL-E calls into verified outcomes using Goal Contracts, Guardian safety checks, and requirement-level evidence.

Comment

Inspiration

AI phone agents can complete a call, but a completed call does not necessarily mean the user's goal was actually achieved.

Future Call AI was built around one principle: provider completion is not verified success.

Future Call AI is powered by Earendel, the evidence-first orchestration and verification layer behind the pipeline: Intent → Goal Contract → Guardian → CALL-E → Evidence → Verify → Result.

The name Earendel is inspired by the extraordinarily distant star whose light reaches us from the deep past — a symbol of carrying human intent across distance and communication barriers. Future Call AI was also designed with accessibility in mind. By allowing an AI agent to handle structured phone interactions on a user's behalf, it can reduce barriers for people who find traditional phone calls difficult, including users with hearing difficulties or language barriers. The current certified demo route is English, while broader multilingual accessibility is part of the project's future direction.

The goal was to create a safer autonomous calling system that can make real CALL-E phone calls while keeping user authority explicit and requiring evidence before claiming success.

What it does

Future Call AI turns a natural-language request into a structured Goal Contract.

Before execution, a Guardian layer checks what the user authorized, which information may be shared, which actions are forbidden, and whether a real call is allowed.

CALL-E then performs the authorized phone call.

After the call, Future Call AI evaluates evidence requirement by requirement. The final outcome can be:

  • VERIFIED_SUCCESS
  • `PARTIAL
  • `UNKNOWN
  • `FAILED

A provider status such as completed is never enough on its own.

The app also includes a simulated mode that uses no CALL-E credits, plus captured production E2E evidence so judges can inspect a verified real-call result without placing another call.

How we built it

The backend is built with Python and FastAPI and integrates the official CALL-E SDK.

The frontend is built with React and Vite.

The execution flow is:

Intent → Goal Contract → Guardian → CALL-E → Evidence → Verify → Result

Safety and reliability features include:

  • explicit user consent before live execution
  • bounded authority and forbidden-data rules
  • E.164 phone-number validation
  • concrete resolution of relative weekday references
  • no invention of missing identity or booking data
  • stable idempotency keys
  • duplicate-call suppression
  • resume-safe polling after refresh or timeout
  • requirement-level evidence verification
  • simulation mode by default

The application is deployed on Render, with automated backend tests and frontend production builds through GitHub Actions.

Challenges

The hardest challenge was distinguishing technical call completion from real-world task success.

A call provider can report a completed interaction even when the requested outcome was not confirmed. We therefore designed the verifier to inspect evidence independently instead of trusting provider completion.

Another challenge was preventing accidental duplicate live calls. We added idempotency, backend duplicate suppression, and frontend job persistence so refreshing the browser does not silently create another paid call.

We also had to handle ambiguous natural-language dates such as "Friday" by resolving them into concrete future dates before sending the task to CALL-E.

Accomplishments

We completed a real production E2E CALL-E test in which a recipient answered, confirmed the requested appointment change, and confirmed that there would be no additional charge.

Future Call AI only returned VERIFIED_SUCCESS` after both requirements had supporting evidence.

The project also provides a safe zero-credit demonstration path for reviewers.

What we learned

The most important lesson was that autonomous agents need more than execution.

They need explicit authority, evidence, contradiction handling, and a clear definition of success.

Reliable autonomy is not simply "the API returned completed." It is proving that the user's actual goal was satisfied.

What's next

Future Call AI could evolve into a reusable verification layer for autonomous phone workflows such as appointment coordination, customer service, reservations, supplier calls, and other bounded real-world tasks.

The next step is to expand Goal Contracts, evidence extraction, and policy controls while preserving the same evidence-first principle.

Built With

Share this project:

Updates

Submission history