A WhatsApp assistant that qualifies property buyers in a short chat, scores each lead warm or cold, and hands warm leads to a sales agent, with a live dashboard.
Property enquiries arrive on WhatsApp. Before a sales agent knows whether a lead is worth a call, someone has to ask the same six questions: name, property type, budget, timeline, location and financing. That work is repetitive and delays the reply to the buyers who are ready to move.
A Node.js service receives WhatsApp messages and makes one Gemini call per turn, returning the reply, the extracted details and the warm or cold score together. Rules like the automation notice and STOP opt-out are enforced in code, not the prompt.
Add a name and number, and the assistant opens the conversation after checking the number is on WhatsApp.
One or two questions per turn, mirroring the buyer's language, collecting name, property type, budget, timeline, location and financing.
Each Gemini call returns the reply, the extracted fields and a warm or cold score as JSON, checked in code before it is trusted.
A fixed disclosure line, STOP opt-out handled before any model call, and a turn limit that defaults to cold.
Warm leads go to the next active agent the moment they qualify, and the buyer is told who will call.
Stat cards and a lead table with each buyer's answers, status and assigned agent, refreshed every five seconds behind an admin login.
A sales admin adds a buyer on the dashboard and the assistant messages them on WhatsApp: first a fixed line saying it is automated, then a short, natural conversation in English, Hindi or Hinglish. It never quotes prices or listings. When it has enough, it scores the lead warm or cold, assigns warm leads to the next agent, and tells the buyer who will follow up.
Architecture · Architecture
Buyer reply message flow · Message Flow
Cold outreach sequence · Sequence
Data model · Data Model
Admin login · Auth Flow
Chose Gemini with a JSON response schema over Free-text replies parsed with regex, or a second extraction call
One call per turn returns the reply, the stop signal, the warm or cold score and the extracted fields together. The code still normalises the score and coerces the done flag before trusting it.
Chose Hardcoded disclosure, STOP before any model call, a turn cap that defaults to cold over Putting every rule in the system prompt
Behaviour that must never fail, such as saying the assistant is automated or honouring an opt-out, does not depend on the model following instructions.
Chose A separate bridge process reached over HTTP, isolated in one transport module over The official Cloud API from day one, or in-process client libraries
The official API needs an approved template for the first outbound message. Two in-process libraries proved unreliable. Isolating the transport means moving to the official API changes one file.
Chose A MongoDB conversation per phone number with embedded history and fields over A separate messages collection or relational tables
Every turn needs the full history for the model call, so a single read and write per turn keeps it simple, and warm leads are one query away.
Chose Next agent derived from the count of assigned conversations over A stored cursor or a queue
Nothing extra to keep in sync. The code accepts that one cycle may misassign when the roster changes, which is fine at demo scale.
Phase 1
Phase 2
Tell us what you're building and we'll show you how we'd approach it.
Consumer apps and creator economyA studio promised its creator client an AI social app in a month with one developer. We scoped the real need and had four roles working eight days after the first call.
Gaming and PC optimizationA full-stack, design-first e-commerce build for a PC optimization brand, with cart, checkout and key delivery, that reached a client-reported 25% CTR in a low-conversion niche.
Gaming and PC optimizationA Discord bot that pulls the best gaming deals from Amazon through its Product Advertising API and posts them to the community every four hours.