PROJECT OVERVIEW
Project Name: AI-Powered Property Recommendation via Contact Request & CRM**
Overview
This project builds an AI-driven recommendation system that responds to user property requests submitted through the **Contact Us** form on a real estate website.
When a user submits a request specifying budget, location, property type, and preferences, the request is:
1. Stored in the CRM as a lead
2. Processed by the backend
3. Passed into a vector search engine
4. Used to retrieve the most relevant property listings
5. Returned to the user via the website chat interface as recommended property links
This system improves lead response speed, relevance, and conversion by providing instant, intelligent property suggestions
MVP(Minimal Viable Product) DEFINITION
MVP Goal
Deliver instant property recommendations based on user specifications from the Contact Us form, while logging the request in the CRM
MVP Scope (Phase 1)
User Flow
1. User clicks Contact Us
2. User submits:
* property type
* location preference
* budget
* specifications
3. Request is:
saved in CRM
sent to recommendation engine
4. User receives:
5–10 recommended property links in chat
Components
- Contact Us form
- Backend API to receive lead
- CRM lead ingestion
- Vector search–based recommendation engine
- Chat response with property cards or links
Recommendation Logic (MVP)
- Hard filters: budget, listing type, location
- Vector similarity: semantic matching of property descriptions
- Re-ranking: closeness to budget + preference match
Out of Scope (For MVP)
- Personalized long-term user profiles
- Payment or booking
- Mobile app
- Analytics dashboard
- Agent assignment logic
MVP Success Criteria
- Recommendation response time < 2 seconds
- At least 70% relevance in manual review
- Works with ≥ 1,000 property listings
- CRM successfully stores every request
ROLE ASSIGNMENT (TEAM OF 4–5)
AI / ML Lead
Responsibilities:
- Define recommendation strategy
- Design embedding & vector search pipeline
- Build Colab notebooks
- Define input/output schemas
- Validate recommendation quality
Ownership:
- Vector index
- Query logic
- Recommendation relevance
Backend Engineer
Responsibilities:
- Contact Us API endpoint
- CRM ingestion
- Recommendation service orchestration
- Chat response handling
Ownership:
- API contracts
- Service reliability
Frontend Engineer
Responsibilities:
- Contact Us form UI
- Chat interface UI
- Rendering recommendation cards/links
Ownership:
- User experience
- API integration
Data / Infra Engineer (Optional but valuable)
Responsibilities:
- Property dataset preparation
- Data cleaning & normalization
- Storage & deployment setup
Ownership:
- Data quality
- Environment stability
Leadership rule:
You own decision clarity, not all the code.
DATASET PREPARATION (YOUR FIRST TECH TASK)
Required Property Dataset Fields
Each property listing must include:
| Field | Description |
| ------------- | ------------------------------ |
| property_id | Unique identifier |
| title | Short property title |
| description | Full description |
| listing_type | sale / rent / shortlet |
| property_type | apartment / duplex / land |
| price | Numeric value |
| location | City / area |
| bedrooms | Integer |
| bathrooms | Integer |
| amenities | List (parking, security, etc.) |
| url | Public listing link |
| status | available / unavailable |
Fields Used for Vector Embeddings
These fields are combined into one semantic text:
- title
- description
- amenities
- location
- property_type
- listing_type
Example combined text:
“3 bedroom furnished apartment in Lekki Phase 1 with parking, 24-hour security, close to the beach.”
Fields Used for Filtering (NOT embeddings)
- price
- listing_type
- location
- bedrooms (if strict)
Dataset Sources
- Existing company database
- Export from CMS
- Scraped listings (if approved)
- Mock dataset (for early MVP)
Output of Dataset Preparation
- Cleaned property table
- Text field for embeddings
- Metadata table for filters
- Ready for FAISS indexing
CRM INTEGRATION (MVP VIEW)
CRM Receives:
- User contact info
- Property request specs
- Timestamp
- Recommendation IDs returned
Why this matters:
- Agents can see what was recommended
- Follow-up becomes context-aware
- Better conversion tracking