The platform. Explained without the marketing layer.

Every WiseAI vertical runs on LiveKit Agents, a production-grade voice AI framework. One agent pool handles all customers across all verticals simultaneously — multi-tenant by design. Each customer's session loads their specific configuration at call time.

We use Deepgram Nova-3 for speech-to-text, with per-vertical keyterms boosted at session start. Cartesia Sonic handles text-to-speech. The voice model does not change; the content, prompts, and routing rules do.

Coordinator and Care agents run Claude Haiku 4.5 with prompt caching — the model that handles the first moments of a call. Caching means the system prompt (thousands of tokens) is loaded once per customer session, not re-transmitted on every turn. Gemini 2.5 Flash handles fallback when primary is unavailable.

No model has access to prior customers' data. Each session is isolated. System prompts are assembled at call start from per-vertical templates + per-customer config + product knowledge loaded from the database.

A Coordinator agent handles intake and routing. A Care agent handles emotionally sensitive conversations. A Sales agent handles product inquiries and lead capture. Each role has its own system prompt, its own tool access, and its own escalation rules.

The Coordinator decides which agent speaks next based on call context. In a funeral home context, the Care agent is primary for 90% of calls. In a church context, the Coordinator handles most general inquiries, the Care agent handles grief and pastoral requests.

Most voice AI platforms have a technology moat. Ours is not just technical — it is pastoral. The HEAR protocol (Hear, Empathize, Advance, Respond) came from Clinical Pastoral Education and Stephen Leader training. It encodes how trained care professionals listen to people in distress.

The protocol is baked into the Care agent's system prompt at the prompt-fragment level — not bolted on as an instruction. Every care-sensitive interaction follows it by default. This is the product behavior that competing platforms cannot copy without lived experience.

Each vertical has its own product knowledge table in the database. Funeral homes load a different knowledge base than churches. At call start, the agent's system prompt is assembled from: the vertical's base template + the customer's specific configuration (services, pricing, staff contacts, chapel names) + the most relevant product knowledge rows.

Knowledge is updateable by the customer through a dashboard. Changes take effect on the next call. No redeployment required.

Alongside voice, every vertical offers a web chatbot built on the Vercel AI SDK 6 with streaming SSE. The same HEAR protocol applies. The same knowledge base. Tool access is gated by plan tier — premium customers can unlock FAQ lookup, appointment capture, and escalation triggers.

The web layer is a Next.js 16 App Router application deployed on Vercel. ISR-rendered public pages load in under 200ms globally. The admin dashboard gives each customer real-time access to call logs, transcripts, and performance metrics.