Projects / AI Voice Agent (Shining Image)
AI Voice Agent (Shining Image)
Answers calls, books jobs, and hands off to humans when needed. Live on Twilio SIP with transcripts to CRM.

AgentsVoiceTwilioLLM
Key metrics
greeting→first response (p50)
320 ms
ASR + NLU + TTS roundtrip
turn latency (p95)
480 ms
caller → reply (steady load)
containment rate
68%
calls fully handled with no human
handoff rate
18%
routed to human with transcript + intent
booking conversion
39%
was 24% pre-launch (inbound calls)
missed-call reduction
↓62%
4-week pre/post
ASR WER (domain)
6.5%
human-checked sample n=150
intent routing accuracy
93%
task chosen matches human label
avg call cost
$0.014 / min
Twilio + LLM + infra blended
avg cost per booked job
$0.23
includes partial human handoffs
uptime (30-day)
99.95%
active monitoring + healthchecks
Methodology: 4-week pre/post cohort; mixed human grading + automated logs; details available on request.
Problem
Missed calls after hours and during peak windows led to lost bookings and slow callbacks. Staff needed a way to capture jobs 24/7 without creating more admin burden.
Approach
- Twilio SIP trunk into Asterisk/FreePBX; DID routing by business hours.
- Low-latency NLU + tool calls: FAQ, pricing estimator, availability lookup, booking creation.
- Human handoff when confidence < threshold or caller asks; warm transfer with screen-pop and transcript.
- Post-call summary + structured notes to CRM; daily call report with failure reasons.
Results
- Missed-call rate reduced by 62% in first 4 weeks; bookings captured 24/7.
- Booking conversion improved from 24% → 39% (+15pp) on inbound calls.
- 68% of calls contained end-to-end with no human agent; handoff on the rest with context.
- Callbacks time dropped from 2h 11m median → 27m due to captured details and auto-ticket.
Stack
Asterisk/FreePBXTwilio SIPWebSockets (streaming ASR/TTS)Node/TypeScriptPythonLLM function callingPostgresRedis
Responsibilities
- Designed call flow & tool APIs (availability, booking, quotes)
- Optimized latency path (streaming ASR/TTS; short prompts; caching)
- Built handoff protocol + transcript packaging to CRM
- Set up monitoring (turn latency, containment, WER, intent accuracy)