Build a CallSphere-Style Multi-Agent for HVAC Dispatch

TL;DR — At $2,500–$5,000 average ticket, missing 5 calls/week is $50–100k/year in lost revenue. Build a 4-specialist HVAC dispatcher (intake, scheduling, parts/quote, emergency) on OpenAI Realtime + ServiceTitan API. Per-minute cost: ~$0.07 vs. 0.30–0.80 for vendor stacks.

What you'll build

A 4-agent HVAC dispatcher: triage → (new-job intake | reschedule | parts/quote | emergency). Books to live ServiceTitan capacity, SMSes confirmations with tech name + ETA window, and pages the on-call for after-hours emergencies.

Prerequisites

1. ServiceTitan API access (Marketplace integration).
2. OpenAI Realtime + Twilio Voice + SMS.
3. Python 3.11+, openai-agents[voice], fastapi.
4. SMS opt-in language reviewed for your state.
5. On-call rotation in Postgres.

Architecture

flowchart TB
  C[Caller] --> TR[Triage]
  TR --> NEW[New Job]
  TR --> RES[Reschedule]
  TR --> PQ[Parts/Quote]
  TR --> EM[Emergency]
  NEW --> ST[(ServiceTitan)]
  RES --> ST
  PQ --> ST
  EM --> PAGE[On-call page]

Step 1 — Triage

```python triage = RealtimeAgent( name="triage", instructions="""You're dispatch for ABC Heating. Identify intent in 2 exchanges: new-job, reschedule, parts/quote, emergency. Emergencies (no heat in winter, gas smell, water leak) hand off to emergency immediately.""", handoffs=[new_job, reschedule, parts_quote, emergency], ) ```

Step 2 — New-job specialist

```python @function_tool async def find_open_slots(zip_code: str, day: str, urgency: str) -> list[dict]: techs = await st.techs.list_by_zip(zip_code, skill_set=["hvac"]) slots = await st.dispatch.aggregate_slots(techs, day, urgency_score=URGENCY_MAP[urgency]) return slots[:5]

@function_tool async def create_job(name: str, phone: str, address: str, slot_id: str, problem: str, system_type: str) -> dict: j = await st.jobs.create(customer={"name": name, "phone": phone, "address": address}, slot_id=slot_id, summary=problem, tags=[system_type]) ref = f"HV-{datetime.now():%Y%m%d}-{j.id:03d}" await sms.send(phone, f"Tech {j.tech_name} arriving {j.eta_window}. Ref {ref}") return {"ref": ref, "job_id": j.id, "eta": j.eta_window} ```

Step 3 — Reschedule specialist

```python @function_tool async def lookup_existing_job(phone: str) -> dict: j = await st.jobs.find_active(phone) if not j: return {"found": False} return {"found": True, "job_id": j.id, "current_slot": j.slot_iso}

@function_tool async def reschedule_job(job_id: str, new_slot_id: str) -> dict: j = await st.jobs.reschedule(job_id, new_slot_id) await sms.send(j.customer.phone, f"Updated to {j.eta_window}. Ref HV-{j.id}") return {"ok": True} ```

Step 4 — Parts / quote

```python @function_tool async def get_diagnostic_quote(zip_code: str, after_hours: bool) -> dict: z = await st.pricing.zone(zip_code) return {"diagnostic_fee": z.diag_fee + (z.after_hours_premium if after_hours else 0)}

@function_tool async def check_part_in_stock(model_number: str, part: str) -> dict: return await st.inventory.find(model_number, part) ```

Step 5 — Emergency

```python @function_tool async def page_oncall(summary: str, address: str, phone: str) -> dict: oncall = await rotation.who_is_oncall() twilio.calls.create(to=oncall.phone, from_=BUSINESS_NUMBER, url=f"https://you/page?summary={summary}") j = await st.jobs.create_emergency(summary=summary, address=address, phone=phone) return {"job_id": j.id, "paged": oncall.name} ```

Step 6 — Twilio bridge

Standard Realtime bridge. Add CSAT IVR ("press 1 if booked correctly") at end-of-call to populate quality metrics.

Step 7 — Reporting

Track: book-rate (booked / total), AHT, average revenue per booking, after-hours emergencies/night, missed-call rate (always 0 with AI). Replace your old IVR dashboard.

Common pitfalls

• Static schedule copies. Always read live capacity — never overbook.
• Generic emergency thresholds. Per-region tuning matters (snowstorm = no heat is critical).
• Tech name privacy. Some companies prefer "your technician" vs. names — config flag.

How CallSphere does this in production

OneRoof Property runs 10 specialists over WebRTC + Pion + NATS — same shape, scaled. Healthcare's FastAPI :8084 ships 14 HIPAA tools. Salon emits GB-YYYYMMDD-### references the way the HV-YYYYMMDD-### pattern above mirrors. 37 agents · 90+ tools · 115+ DB tables · 6 verticals. Pricing $149/$499/$1499 flat — no $0.30–$0.80/min surprise. Start a 14-day trial or check /pricing.

FAQ

ServiceTitan tier needed? "Marketing Pro" or higher for full API.

TCPA? Inbound is fine; outbound recall to cell phones needs PEWC.

After-hours pricing? Tool returns the surcharge so the AI can quote correctly.

Multi-trade (HVAC + plumbing + electric)? One trade parameter on every tool.

Voice tone? Use ElevenLabs for warmer voice for residential calls.

Sources

• ServiceTitan AI voice
• HVAC voice agent guide
• OpenAI Realtime
• /pricing