I build agentic voice AI both ways: chained STT→LLM→TTS for control and tooling/MCP, or speech-to-speech (Hume, Cartesia, Grok Voice) when latency is the priority - on LiveKit, Retell, and Twilio.
A voice agent that cannot transfer, write to the CRM, or survive a noisy job site is a toy. I treat telephony as product work: numbers, recording consent, interruption handling, and a human handoff when the caller is angry or the job is an emergency.
I set up agentic voice AI on both architectures. Speech-to-speech (Hume, Cartesia, Grok Voice) is more latency-friendly for natural phone talk, but tooling and MCP support are harder. A chained STT→LLM→TTS pipeline gives more control over each layer, but each stage needs fine-tuning. Sometimes Retell is right because ops edits scripts themselves; sometimes LiveKit plus the right pipeline is the product.
After the hang-up, the call has to land somewhere: GoHighLevel, eSimply, or an internal CRM. Speed-to-lead for realtors and missed-call recovery for HVAC are the jobs this service is built around.
Who it is for
- HVAC, roofing, and home-service companies missing peak-season calls
- Realtors who need form-to-phone speed-to-lead
- SaaS products replacing cascaded STT-LLM-TTS with lower latency
What you get
Telephony that answers
Twilio inbound and outbound, Retell or custom LiveKit rooms, and transfer rules to a human when it matters.
Latency you can feel
Speech-to-speech (Hume, Cartesia, Grok Voice) when latency wins. Chained STT→LLM→TTS when you need per-layer control, tooling, and MCP - with fine-tuning at each stage.
CRM after the hang-up
Outcomes written to GHL, eSimply, or an internal CRM so the next human is not starting from zero.
How I work
Call types
List the top five intents and the emergency / escalate path before recording a voice.
Pilot DID
One number, one flow, recordings reviewed weekly.
Outbound
Add a campaign (duct cleaning, inspections, seller leads) once inbound is stable.
What this looked like for Lehigh HVAC
Custom CRM as system of record - customers, jobs, follow-ups, campaign lists, call outcomes. Retell front desk for scheduling, FAQs, intake, and urgent escalations. Outbound Retell for duct cleaning: pitch, objections, leads back into CRM. Make.com ties voice events to the office workflow.
Proof
- Lehigh HVAC: CRM plus Retell front desk plus duct-cleaning outbound
- Use Cortana: LiveKit cascade migrated to Hume S2S, thousands of monthly calls
- Realtor builds: Zapier to GHL or eSimply plus Retell speed-to-lead
Case studies
- Lehigh HVAC CRM and Retell
- LiveKit to Hume at Use Cortana
- Open-source LiveKit cascade demo
- VoiceAI-Backend on GitHub
- VoiceAI-Frontend on GitHub
- Use Cortana agent studio + walkthroughs
- Use Cortana walkthrough Part 2
- Real estate lead plus Retell
Stack and docs
Related
Common questions
Retell or custom LiveKit?
Retell when operators must edit scripts themselves. LiveKit plus Twilio when you need a productized studio, multi-tenant config, or a custom chained / speech-to-speech pipeline.
Chained pipeline or speech-to-speech?
STS (Hume, Cartesia, Grok Voice) is more latency-friendly; tooling and MCP are harder. Chained STT→LLM→TTS gives more control per layer and stronger tool wiring, but each stage needs fine-tuning. Pick from latency vs control - I ship both.
Will this replace my receptionist?
It covers repetitive rings and after-hours. Emergencies and messy diagnostics still escalate. The win is missed-call recovery, not firing the office.
How fast can a calling agent go live?
A single-flow MVP is usually weeks if the CRM and number are ready. Multi-agent platforms take longer and start with a discovery call.
Can the agent transfer to a human on my cell or the office line?
Yes. Transfer rules are part of the build: emergencies, angry callers, or anything the script cannot finish. The agent is the front desk, not a wall.
Do I need a CRM first?
You need somewhere the outcome lands: GoHighLevel, eSimply, or an internal CRM. If you have none, I can ship a thin one (as with Lehigh HVAC) or wire the agent to a spreadsheet as a stopgap, then replace it.
What about call recording and consent?
Recording and disclosure follow your jurisdiction and carrier rules. I build the prompt and the telephony flags you need. Legal sign-off stays with you.
Does it work after hours and on weekends?
That is a main use. After-hours and overflow rings are where HVAC and realtor agents earn their keep. Hours and holidays are config, not a code change every time.
Which languages can it speak?
English is the default I ship. Other languages depend on the STT/TTS or speech-to-speech vendor. Tell me the language on the intake call before I pick a stack.
What happens if Twilio or the model is down?
Calls should fail over to a human number or voicemail, not dead air. I treat provider outages as a product path, not an afterthought.