Stella: a guest-inquiry assistant that routes by value
I designed and prototyped an assistant that answers routine questions, books small tables and sends high-value requests to a manager. The demo shows each routing decision.
- Question
- Which guest requests can an assistant close alone, and which must go to a person because the value or the risk is high?
- Data
- A knowledge base of 13 chunks (menu, hours, policies) and a set of routing rules. The demo uses sample content only.
- Method
- Answer facts from the knowledge base only. Book parties of 1 to 5 after the name, phone, party size, date and time are known. Send parties of 6 or more, private rooms, corporate dinners, catering, buyouts and requests over $500 to a manager.
- ResultPrototype
- A working text demo with a rule router and a routing log. The voice version ran on Retell AI with n8n, Pinecone, Airtable and Twilio, and went live as the Autostrada phone line.
One design rule
The language model never sends the manager text. A fixed function sends it. Thus the model cannot claim that a text went out when it did not.
From prototype to live phone line
I designed and prototyped Stella first. Its voice front end, on Retell AI, then went live as the Autostrada phone line. The web demo below is a later text rebuild with sample data.
The live phone line: Retell call routing
Real data
- Guest calls. A guest calls the restaurant.
- Retell voice agent. Answers, asks the intent and routes the call.
- Hours, location and menu. Answered on the call.
- Reservation. Booked or changed.
- Event or large group. Details taken for follow-up.
- Other or complex. Passed to staff, then transferred to the host or manager.
The voice front end of Stella went live as the restaurant’s phone line. I set up the agent, its answers and the routing rules.
The demo is a later text rebuild with sample data. The menu, prices, staff names and phone numbers are sample content. Bookings and manager texts are simulated. The demo is not a production system.
More work
- Stay for the exhale
I turned a yoga teacher's classes into a membership product. I did the strategy, the design, the build and the operations.
- A bank reconciliation that explains each difference
A working sample of the reconciliation method from my years as Operations & Finance Lead. Excel formulas match 119 bank lines to 121 ledger entries, explain each difference and reconcile to 0.00.
- SQL case studies: revenue share, donor segments and a 3NF schema
Three SQL studies on small course datasets. Each one starts with a business question. Every figure is query output that you can run again.