AI Voice Collections Agent
Real-time collection conversations with deterministic policy controls.
Checking backend…
Streaming voiceMic → streaming STT → end-of-turn → policy → LLM/templates → streaming TTS → speaker.
Barge-inCaller speech cancels agent audio mid-sentence; stale audio is dropped by generation id.
Policy guardrailsIdentity-before-disclosure, payment envelope, stop-contact, calling hours — deterministic and audited.
Promise-to-payOnly confirmed after an explicit “yes” to a read-back that was actually heard.
Japanese / EnglishExplicit language selection; bilingual prompts, templates, number and date parsing.
PSTN adapterTwilio Voice + Media Streams behind a TelephonyProvider; disabled unless configured.
Evaluation harness30 regression scenarios with deterministic invariants; optional LLM-as-judge.
Audit trailEvery policy decision, state change, barge-in and provider failure is reconstructable.
Core design principle
Language models handle language. Application code owns authority.
The model may interpret the caller and propose typed actions. It cannot verify identity, disclose the balance, approve a payment plan, set stop-contact, confirm a promise or perform a transfer — those are deterministic state transitions in the conversation controller, each gated by the policy engine and written to the audit trail.
Two-minute reviewer walkthrough
- Open the voice demo, pick English or 日本語 and scenario A.
- Verify identity with the synthetic date of birth shown on the card.
- Offer ¥30,000 in two weeks, then confirm — watch the promise-to-pay panel.
- Try scenario C (60-day extension) and watch the policy engine block it.
- Interrupt the agent mid-sentence (scenario F) and read the barge-in latency.
- End the session and open its audit trail; then run the evaluation suite.