RealGuard
AI automation and conversational workflows for real estate. WhatsApp conversational input → Groq parameter extraction → Java business logic → RERA lookup → in-app EMI calculation → CRM persistence → broker workflow. The LLM extracts parameters while application logic handles authoritative calculations.
Executive Summary
RealGuard is a conversational AI automation prototype engineered to qualify real estate leads over WhatsApp while preventing AI calculation errors through deterministic backend boundaries.
Project Context & Core Attributes
- Problem: Real estate brokers lose prospective buyers due to slow manual response times, while generic AI chatbots frequently hallucinate property prices, interest rates, and loan amortization numbers.
- Intended User: Real estate brokers, property consultants, and agencies managing inbound buyer inquiries via WhatsApp.
- My Role: End-to-end architecture and backend implementation (conversational intake pipeline, Groq LLM extraction, deterministic calculation engine, and CRM lead persistence).
- Solution: A decoupled workflow where Groq LLM extracts unstructured buyer criteria into strict JSON, while Spring Boot deterministically calculates loan EMIs, queries RERA registrations, and pushes clean lead records to CRM.
- Verified Status: Working architecture prototype with verified Twilio webhook parsing, deterministic calculation logic, and open-source GitHub repository.
- Evidence: Open-source repository at github.com/shivam-shukla888/RealGuard.
Architecture & Conversational Flow
The system enforces strict separation between conversational interpretation and authoritative domain calculations:
- WhatsApp Webhook Intake: Prospective buyers message through WhatsApp via Twilio webhooks.
- Conversational Parameter Extraction: Groq API (Llama 3) parses unstructured chat to extract budget, preferred locality, BHK requirements, and timeframe into structured JSON parameters.
- Java Business Logic Routing: Spring Boot services validate incoming parameters against application boundaries and route to specific sub-modules.
- RERA Registry Lookup: Queries registered project records to cross-reference developer legitimacy against official regulatory records.
- Deterministic EMI Calculation: Computes loan interest, tenure amortization, and monthly EMI figures strictly within deterministic application code — the LLM never performs financial calculations.
- CRM Persistence & Broker Notification: Stores buyer records and interaction history in MySQL via Hibernate ORM and dispatches structured lead summaries to real estate brokers.
Key Engineering Lessons & Defensive Boundaries
- AI for Extraction, Code for Math: Language models interpret conversational nuance, but authoritative financial calculations and compliance rules belong strictly in verified application code.
- Webhook Integrity: Spring Boot verifies Twilio webhook signatures before processing messages to prevent forged broker dispatch events.
- Structured Lead Qualification: Normalizes free-form text conversations into standardized relational CRM records without requiring rigid UI forms.
Implementation Stack
Engineered with Java 17, Spring Boot, MySQL, Hibernate ORM, Twilio WhatsApp API, and Groq API.
Questions about this project?
Feel free to reach out if you'd like to talk about the tech stack, implementation, or a similar project.