No invented policy
Open questions become messages or owner-attention items instead of confident guesses.
Case study · AI reception & operations
Brandi,
Wonkeeland’s AI front desk
A receptionist built for the actual policies, schedule, customers, service area, and working day of an owner-operated repair business.
BRANDIAI Front Desk
VoiceCalendarCRMSMSBillingClientWonkeeland Appliance Repair
SystemAI reception + operations
DeploymentDedicated self-hosted appliance
Product evolutionFoundation for River™ Reception
Leon cannot always stop an appliance diagnosis to answer the telephone. But every caller still needs accurate information, a useful next step, and a receptionist that follows the business’s actual rules. Brandi was built to close that gap without pretending uncertain AI output should control pricing, identity, availability, or customer records.
01
The challenge
A phone call can touch the whole business.
A caller may need service-area qualification, appliance eligibility, pricing context, appointment availability, rescheduling, cancellation, a message, or a human. The correct response can depend on travel time, calendar conflicts, customer verification, existing appointments, and business policies.
A convincing voice alone would not solve that problem. The receptionist needed a dependable system behind the conversation—and clear limits around what the model was allowed to decide.
02
The architecture
The model speaks. Code owns the facts.
Incoming callTelnyx
→Live conversationOpenAI Realtime
→Deterministic toolsPolicies & authorization
→Business systemsCalendar · CRM · Routes · SMS
Brandi uses real-time speech and language understanding for the conversation, while deterministic application code controls service eligibility, pricing facts, identity verification, scheduling state, and database changes. Caller ID can support a friendly greeting, but it never authorizes private disclosure.
03
What we built
More than an answering bot.
- Real-time inbound voiceAuthenticated bidirectional phone audio, interruption handling, recovery behavior, tool calls, and transcript persistence.
- Deterministic schedulingAvailability, conflict checks, booking, rescheduling, cancellation, time blocking, and idempotent writes.
- Service-area intelligenceTraffic-aware drive-time checks and explicit handling for business-approved exceptions.
- Customer and job recordsPersistent customer, address, appliance, job, interaction, message, verification, audit, and attention records.
- Messaging workflowsInbound and outbound SMS/MMS, delivery events, consent commands, deduplication, confirmations, and owner notifications.
- Owner consoleAuthenticated mobile access to schedule, customers, calls, messages, attention items, system status, and appointment actions.
- Billing systemDraft invoices, tax and exemption handling, payments, refunds, secure customer pages, and PDF documents.
- Operations and recoveryHealth/readiness separation, watchdogs, encrypted off-site backups, deployment receipts, and documented rollback.
04
Guardrails
Useful automation requires deliberate restraint.
Authorization in code
Private information and owner actions are gated independently of conversational persuasion.
Auditable behavior
Effective configuration, prompts, models, voices, tool outcomes, and operational events are recorded.
Safe failure
Provider errors, silence, interruptions, duplicates, and incomplete workflows have explicit recovery paths.
Verified engineering evidence
Deployed, monitored, and built to be tested.
The public health endpoint is live, and the project includes automated unit, integration, scenario, browser, provider-contract, security, and persistence coverage. Tests use fake transports rather than placing unauthorized live calls or modifying provider accounts.
Public healthLiveVerified August 30, 2026
Data foundation20Versioned database migrations
RuntimeSelf-hostedDedicated hardware + PostgreSQL
RecoveryDocumentedBackup, preflight, receipt, and rollback gates
05
The product outcome
Brandi became the proving ground for River Reception.
Brandi demonstrates that an AI receptionist should not be sold as a personality prompt attached to a telephone number. The value is the configurable operational layer behind her: approved knowledge, deterministic tools, business integrations, escalation, review, and ongoing improvement.
That reusable foundation is evolving into River™ Reception, while each business can still receive its own receptionist name, voice, knowledge, call flows, and integrations.
Customer-volume and revenue-impact claims will be added only after enough monitored production traffic exists to support them.
Calls are part of the workflow