Skip to content
TURBO DIGITAL DEVELOPMENT
ENES
Home/Case studies/Hybrid AI receptionist

Case study · Staging platform

Smithfield AI: dental reception with reliable messaging and staff control.

Smithfield AI helps dental teams organize SMS inquiries, capture appointment preferences and prepare follow-up. A hybrid agent uses OpenAI for language interpretation and drafting, while deterministic rules control scheduling, opt-outs, escalation and response validation. Staff retain responsibility for confirming appointments.

Built with Node.js, a browser-based staff dashboard and encrypted PostgreSQL persistence, the platform includes tenant-isolated clinic workspaces, durable messaging and NexHealth sandbox integration. The current implementation is in staging for client evaluation; live clinic deployment remains subject to integration, infrastructure and operational approval.

Node.jsHybrid AITwilio SMSPostgreSQLTenant isolationOpenAPI 3.1GitHub Actions
Explore the servicePlan a controlled pilot

AI with limits

Rules constrain model output

Opt-outs and sensitive requests take precedence. Invalid or unsafe generated replies fall back to approved wording.

Durable processing

Accepted is not delivered

Persisted inbox and outbox records separate message receipt, processing and carrier delivery.

Clinic boundaries

Separate workspaces and permissions

Tenant-scoped sessions, data queries and provider bindings protect each location's workflow.

How it works

From SMS to staff follow-up.

Every request follows a clear path. Staff stay in control of appointment confirmation.

  1. 01 / Receive

    Receive the message

    Validate the provider webhook and save the incoming SMS in the clinic's workspace.

  2. 02 / Interpret

    Understand the request

    Identify intent, enforce opt-outs and route sensitive questions to staff. Ask for clarification when the request is unclear.

  3. 03 / Offer

    Offer approved times

    Use current, staff-published availability or capture another preferred time. The assistant never invents slots or confirms a booking.

  4. 04 / Hand off

    Prepare staff follow-up

    Bring the conversation, preferred time and suggested next step into one dashboard. Staff review replies and confirm appointments.

Engineering decisions

Handling retries, concurrent work and clinic data.

The application uses a modular monolith with replaceable providers. Its PostgreSQL path makes message processing durable without requiring a separate microservice for every function.

Tenant isolation

Scope every operation to a clinic

Staff sessions, service-account scopes, data queries and provider credentials are tenant-bound. Staging locations run in separate workspaces; this is not yet a shared multi-tenant production server.

Concurrency

Preserve conversation order

Database row locks and expiring worker leases coordinate background processing. Messages for one phone are handled in order, while provider calls run outside the state transaction.

Idempotency

Recognize repeated events and commands

Persisted webhook and command identities prevent duplicate processing. Unchanged retries return the retained result; changed content using the same identity is rejected.

Failure recovery

Do not blindly resend uncertain messages

Failed inbound jobs back off before entering a retained dead state. Ambiguous outbound sends pause for reconciliation, and signed delivery callbacks update their status.

Data protection

Encrypt retained application payloads

PostgreSQL payloads use AES-256-GCM encryption bound to tenant and record identity. Role checks, scoped service accounts, signed webhooks and tenant rate limits protect API access.

Outreach preview

Review openings, audience and wording

Staff can inspect sandbox openings, screened patient lists and grounded outreach drafts. Saving a preview sends no patient messages; bulk outreach remains disabled.

Testing and delivery

Check behavior before a change reaches a clinic.

Automated checks cover application logic, agent behavior, tenant boundaries and durable database processing. Human review remains part of the release process.

Continuous integration

Node 22 + 24 and PostgreSQL checks

GitHub Actions runs syntax and regression tests, secret scanning, dependency audits, PostgreSQL integration tests and an authenticated staging smoke test. CI does not automatically deploy or send live SMS.

API contracts

Versioned endpoints and scoped access

OpenAPI 3.1 and an interactive API reference document the integration boundary. Tests exercise authorization, cross-tenant rejection, scheduling rules and messaging retries.

Agent evaluation

Repeatable cases and reviewed conversations

Fixtures check ambiguous replies, rejected slots, opt-outs and unsafe model output. Staff score retained conversations for accuracy, tone and booking safety before approving presentation evidence.

Evaluate a controlled pilot before live deployment.

The current platform supports staging evaluation and synthetic NexHealth appointments, not live practice-system bookings. Production launch requires approved clinic access, outreach consent, managed infrastructure, verified scheduling behavior and agreed staff procedures. This case study does not claim production booking results or a healthcare compliance certification.

Plan the pilotView the performance platform
Project brief