Fraud Ops Caller

• 

  Both-party evidence in one investigation. Separate complainant/respondent rails; never quote the counterpart.

• 

  Complainant outcome writes back: next action + quotes. “I will not put card numbers in there” — no PAN on the call.

• 

  Both outcomes side by side, statements isolated, party-specific share links only.

• 

  Plan-first before dial: NVC frame, dispositions, quiet-hours/attempt rails. Destinations masked on the call plan.

Fraud desks still chase both-party evidence by hand — and one unsafe dial becomes a compliance mess.

Fraud Ops Caller is a desk, not a recovery engine. A pack-fired case opens → you see the plan first (what will be asked, what’s forbidden) → confirm demo or live → CALL-E places the call → a disposition JSON writes back. Humans only step in on the rails.

Demo lead: safety-bounded evidence collection (complainant + respondent, side by side).

• Hard rails: no PAN/CVV on the call, quiet hours, attempt caps, hostility two-strike, fail-closed handoff
• Live is opt-in: operator secret + grant bound to the approved number; CALL-E traffic pinned to production HTTPS; phones/transcripts masked in API + UI
• Ambiguous live outcomes halt for reconciliation — no silent retry into a conflicting submit
• Same runtime also covers KYC chase, collections, and merchant outreach (same rails)

Not claiming: hosted ML recovery, consortium scores, chargeback CRM, or “CALL-E invents the case.”

Links

• Awesome PR: https://github.com/CALLE-AI/awesome-phone-call-agents/pull/395
• Upstream: https://github.com/pamu512/fraud-ops-caller

Built with: Next.js, TypeScript, CALL-E SDK, NVC-framed task prompts

Built With

• call-e
• next.js
• nvc
• typescript