02 · Vertical voice AI product
Anruvo
An AI call-intake platform for the German Mittelstand that turns conversations into structured, rule-validated operational cases instead of unowned transcripts.
First vertical
SHK service businesses in Germany
Core decision
Conversation is probabilistic; operations stay deterministic
Status
Functional simulator, workspace and SIP entrypoint
The problem
A missed call in a service business is often a missed order or an incomplete emergency report. A generic voice bot can answer, but a transcript alone does not tell the team whether gas odor changes the safety path, whether a heating failure is urgent or which callback rule applies.
The architectural decision
Anruvo separates conversational intelligence from operational authority. Voice AI handles disclosure, dialogue and structured extraction. A deterministic intake engine evaluates safety, service area, opening hours, customer status, vulnerability and priority, then returns the rule trace, missing fields and next step.
From call to operational case
Receive the call
A signed OpenAI Realtime SIP webhook accepts only incoming-call events and configures a German audio session with AI disclosure.
Extract structure
The conversation becomes typed fields such as caller, location, asset, issue, urgency and safety signals.
Apply business rules
The SHK engine evaluates gas odor, water leak, heating failure, service area, hours and routing without asking the model to improvise policy.
Hand off to people
A tenant-isolated workspace presents the complete case, priority, rule trace and recommended next action to the team.
What the repository proves
A real product boundary
The public cockpit is a demo surface; the protected application reads organizations and cases from Supabase with tenant isolation. Marketing, product preview and authenticated workspace are deliberately separated.
Deterministic vertical logic
The same intake engine powers the simulator and POST /api/intake. It always returns missing fields and the decision path, making behaviour testable before a voice adapter invokes it.
Secure voice entry
The Realtime webhook verifies signatures, accepts a narrow event type and configures an explicit disclosure. Recording and durable transcript storage are not silently enabled.
Controlled pilot funnel
Pilot requests are validated on client and server, protected by a honeypot and rate limiting, and degrade to a clear fallback when email delivery is not configured.
Reliability boundaries
Operational priority comes from explicit business rules, not from free-form model reasoning.
Every intake result includes the rule trace and missing required fields.
Customer data is scoped to the authenticated organization in the protected workspace.
The product does not present demo data, recordings or incomplete voice infrastructure as a live customer deployment.
Current implementation boundary
The signed SIP entrypoint and deterministic intake engine exist. A durable sideband controller that connects the live Realtime conversation to tool execution and persisted case creation is the next milestone. Audio recording and long-term transcript storage are not implemented.
Why Anruvo belongs in the top four
Anruvo shows product judgment as much as engineering. It narrows voice AI to a costly, concrete workflow, makes the business rules inspectable and gives human teams a usable operational object. That is a stronger commercial story than a general-purpose phone bot.