Voice AI Receptionist Platform

Value Delivered — 2-Day Production Build

noboxAI (Jonathan Gatlit) 2026-04-04 — 2026-04-06 LIVE — 4 Tenants

Executive Summary

In two days, we built a production-ready multi-tenant voice AI receptionist platform serving 4 businesses on a single bare-metal GPU server. The system handles inbound phone calls via Twilio SIP, qualifies callers through a 3-gate conversational flow, offers connect-now or scheduled callback, and routes to human agents — all self-hosted with zero cloud AI dependency.

4
Tenants Live
100%
Conv Quality
4%
Re-Ask Rate
293ms
Avg Latency
104
Tests
32
Call Summaries
12.5K
Lines Added
35
Commits

Tenants in Production

GDS Atlanta
Document Shredding
(404) 882-9866
CRU Insurance
Commercial + HNW Insurance
(843) 396-6212
noboxAI
AI Consulting
(404) 620-3288
North Group RE
Yana Dorobantu — Real Estate
(404) 948-5255

What We Built

Live

Multi-Tenant Voice Agent

Single container serves 4 businesses. DID → Redis → per-tenant prompt, voice, greeting, transfer target. New tenants onboard in minutes via automated script.

Live

4-Layer Transfer Architecture

Compensates for Qwen3-8B's ~0% native tool call rate. Caller demand, caller confirm, agent text fallback — 100% transfer reliability.

Live

Gate Redesign (60% → 4% Re-Ask)

SlotTracker extracts 13 fields from caller speech. PolicyEngine handles transfer/callback/hangup deterministically. LLM owns conversation only.

New

Audio Noise Defense

4-layer stack: DTLN noise suppression → Silero VAD → STT confidence filter → interruption handling. Reduced noise interruptions from 6+/call to 1-2.

New

Caller Info Capture

Agent asks name, business name, address naturally after qualifying. Call summary JSON persisted with pre-formatted SMS text.

New

Scorecard & CI

25 scenarios, 104 unit tests. Automated baseline testing across all tenants on every prompt/model change.

System Architecture

Every call flows through deterministic defense layers (blue) before reaching the non-deterministic LLM (amber). The deterministic path runs synchronously and always wins the race against the async LLM.

End-to-End Call Pipeline
Transfer Decision — Deterministic Wins the Race

Production Metrics

MetricValueNotes
Tenants live4GDS, CRU, noboxAI, North Group RE
Service re-ask rate4% (was 60%)Gate 1 reframe + SlotTracker
Conversation quality100%25 scenarios, all natural
Transfer reliability100%4-layer fallback architecture
Native tool call rate8%Fallback compensates for 92%
Avg response latency293msQwen3-8B + /no_think
Call summaries persisted32JSON with SMS text pre-formatted
Unit tests104Was 47 at session start
Scorecard scenarios25Across all 4 tenants
Reliability by Component

Delivery Timeline

2-Day Build — Feature Delivery

What's Next

ItemPriorityStatusDependency
Toll-free A2P verificationHighPending TwilioUnblocks SMS delivery
STT confidence dataHighDesignedCustom Whisper wrapper
CRU renewal re-ask fixMediumKnown issuePrompt iteration
Production call volumeMediumAccumulatingTime (50+ calls)
Log persistenceMediumPlannedDocker logging driver
Transfer classifierLowDesigned100+ call summaries
Tool-tuned SLM evalWatchFutureQwen3.1 / Llama 4

Your Input Needed

1. SMS Delivery — Approve sending SMS to transfer targets on every call?
Recommended: Yes. When toll-free verification completes, SMS context handoff activates automatically. Safe recipient list already configured. Pre-formatted text in every call summary JSON.
2. Production Monitoring — How should call summaries be surfaced?
Recommended: Dashboard. Call summary JSONs accumulate but aren't surfaced. Options: daily email digest, real-time dashboard, or on-demand review.
3. Noise Threshold — Acceptable noise level per tenant?
Recommended: Per-tenant tuning. VAD at 0.38 is balanced globally. Noisy environments (GDS callers from construction sites) may need higher threshold. Quiet environments (North Group RE) can be more sensitive.
4. New Tenant Pipeline — Next tenants to onboard?
Ready to execute. Onboarding script automates all 7 steps: Twilio number → LiveKit SIP → prompt generation → Redis → deploy. Single command, ~5 minutes per tenant.