Back to work

Applied AI · Booking

SoffIA

I built a WhatsApp booking agent for clinics. The model interprets requests; the backend checks availability and scheduling rules before saving a booking in PostgreSQL.

Dashboard in a dental demo environment. Image edited to conceal personal data.
Period
2026
Role
AI Engineer
Team and scope
Own product: application, orchestration and connection to booking operations.
Status
Used at an aesthetic clinic in 2026.
  • TypeScript
  • Next.js
  • PostgreSQL
  • Vercel AI SDK
  • Redis

WhatsApp support and bookings

I built SoffIA as a CoreOS product for handling enquiries and bookings in clinics. It was used at an aesthetic clinic in 2026. The work connected WhatsApp conversations to a booking calendar and a multi-tenant operations panel.

What I built

I implemented the application in TypeScript, Next.js and PostgreSQL. The orchestration enables tools according to the conversation state: checking availability, proposing a time and creating a booking each have different conditions.

The backend validates scheduling rules before saving an appointment. Conflicts over the same slot and retries require database-level control, with conditional writes and locks depending on the operation. The state machine orders the dialogue; authorization and persistence are the server's responsibility.

I used reactive-fsm (version 1.4.0) to control tool availability.

An overly strict validator

The worker rejected a historical tenant identifier because its UUID validation was too strict. I corrected the validation to preserve the identifier and accept its canonical representation, then added regression checks for the formats involved.

Results at one clinic

The clinic's internal record reports response time dropping from 47 minutes to under 10 seconds and no-shows from 25% to 12%.

Try booking control

Try an available slot, an occupied one and a repeat request. The example runs in your browser with fictional data and no backend.

One booking, two outcomes.

An illustrative reconstruction with fictional data. Runs in your browser.

Request

Flow state

  1. Request
  2. Calendar rule
  3. Result
Choose a slot and run the request.

State is temporary and resets when you reload the page.

Related code and writing

Contact

I’m looking for backend or applied AI engineering roles, working on integrations and enterprise software. Remote from Peru (UTC-5).