AI Front Door

Private / PM Only

TATEKANDesk

TATEKANDesk is the AI voice help desk of TATEKANOS. It provides 24/7 first-call intake for general usage questions, basic app guidance, and initial trouble reports across Core, Toriteki, Mitsumori, dMemo, and Time.

0. Current Status

Implementation

The canonical bridge implementation was completed and approved on 2026-04-05. Real phone conversation quality was confirmed in integration testing.

Production Readiness

The bridge is implemented, but final always-on service enablement and final inbound audit verification remain as next tasks.

1. Service Role

Position

  • Front-end voice entry point of the TATEKAN ecosystem
  • Not a replacement for each business app
  • Acts as intake, routing, and escalation layer

Accepted Topics

  • General operating guidance for TATEKANOS
  • Basic usage questions for Toriteki, dMemo, Mitsumori, and Time
  • Initial trouble reports before human follow-up

2. Architecture

PBX Layer

Asterisk 20 on tatekan-pbx connects directly to the Rakuten SIP network and forwards calls over AudioSocket.

AI Layer

scripts/gemini_bridge.js bridges Asterisk audio to Gemini 3.1 Flash Live and returns low-latency audio responses.

Knowledge Layer

The bridge loads active TATEKANDesk knowledge from knowledge_registry.json and the general voice help desk knowledge base.

Audit Layer

Structured JSON logs record caller ID, answer status, source IDs, and session end reasons for audit and operations review.

3. Implemented Features

Conversation

  • AI greeting trigger at session start
  • Bidirectional audio conversation over phone
  • Idle timeout at 30 seconds and hard session limit at 180 seconds

Security

  • Secret Manager based API key retrieval
  • No plaintext secrets on disk in the final canonical design
  • Asterisk management interfaces disabled and outbound calling structurally blocked

Operations

  • Registry-based document ingestion for allowed desk knowledge
  • Tag-based resolution and escalation audit status
  • systemd unit prepared for persistent service operation

4. Human Escalation Policy

  • AI first-call intake is available 24/7
  • Human callback guidance is for weekdays 9:00-18:00
  • Unresolved or sensitive cases are routed to human follow-up instead of forcing AI-only closure

5. References