Text-Me Assistant -> Da Vinci Gelato

The goal of this project is to reduce constant staff interruptions by replacing direct texts with a virtual assistant. This AI-powered SMS assistant answers common questions and escalates anything it can’t handle to a manager, learning from their response.

How It Works

1. User texts the business number (we use Twilio).
2. Assistant checks:
   • Hardcoded qa.json prompt (QA_TEXT)
   • Previously learned answers from Supabase qa_pairs table
3. If a match is found:
   • Assistant replies with the appropriate answer
4. If no match:
   • Assistant responds: "Let me forward this to a manager."
   • Manager is alerted via SMS
5. When the manager replies:
   • Their response is forwarded to the original user
   • The Q&A pair is saved for future use

Assistant Behavior

The assistant is instructed to only answer using approved information:

"You must answer employee questions ONLY using the exact information provided... If you cannot find the answer in the information below, respond with exactly: Let me forward this to a manager."

Tech Stack

• OpenAI GPT-3.5 Turbo –> Language model for generating responses
• Twilio –> Handles SMS send/receive
• Supabase –> Stores Q&A pairs and escalation logs
• AWS SAM –> For packaging and deploying serverless functions
• Node.js –> Backend logic

Some Next Steps + Improvements

• Modularize all the core logic into helper files like (db.js, ai.js, sms.js, utils.js)
• Clean up the redundant code in index.mjs and centralize the Twilio/Supabase client setup
• Handle concurrent incoming messages (race conditions)
• Create a dashboard for the manager to review/edit answers