Avoca

Vertical-first conversational AI: Avoca appears to be building a voice + messaging stack purpose-built for home services (HVAC, plumbing, electrical). That implies domain-specialized ASR/NLU models, an intents/slots ontology tuned to booking, dispatch, estimates, and common service-speak (e.g., equipment, symptoms, urgency). Vertical specialization reduces generic LLM hallucination risk and enables precise routing/booking behavior that general contact-center bots struggle with.

Product+services engineering loop (Forward-Deployed Engineers as a core technical pattern): Rather than pure self-serve SaaS, they embed engineers with customers to configure dialogue flows, integrations, edge connectors, and local business rules. That's an unusual architectural choice: the deployed service is not only a multi-tenant platform but a product that intentionally expects per-site operational configuration and model adaptation as a first-class component.

Hybrid low-latency inference architecture implied by 'answer every call' promise: To consistently answer and transact on calls in real time they likely combine streaming ASR, a local/edge inference tier (or prioritized low-latency cloud endpoints) for real-time routing and slot filling, plus asynchronous/centralized LLMs for complex reasoning, coaching insights, and campaign content generation. The split-path (real-time vs async) design is non-trivial and uncommon in simpler chatbot stacks.

Deep operational integration with dispatch and scheduling systems: 'Book jobs 24/7' and 'fill your board' implies two-way transactional integrations (read availability, write appointments, handle conflicts, confirm technicians). This requires idempotent write operations, optimistic locking or distributed transaction handling across third-party field-service CRMs (ServiceTitan/Jobber-like systems) — a hidden complexity often glossed over in marketing.

Call scoring + coaching as a real-time and offline ML pipeline: Surfacing 'missed opportunities' and giving managers 'real-time visibility' indicates automated QA and inference pipelines that (a) process audio→transcript→features, (b) run classification/ranking models for opportunities, and (c) generate actionable coaching items. That requires annotated datasets, active learning workflows, and near-real-time feature stores — an operational ML complexity many startups underestimate.