AI enabled
real estate marketplace

Abto Software has cooperated with a US-based startup setting some ambitious goals in the housing industry.

The project was aimed at delivering a custom, cross-platform solution with rich, easy-to-navigate functionality, enhanced with computational technology to provide higher competitiveness on the real-estate market.

Speaking about the project’s main goals:

• Software development
  At the first stage, our team was focused at designing a marketplace with intuitive property search based on specific criteria and filters.

• ML training and implementation
  At the next stages, our engineers were working towards implementing machine learning for streamlined property matching and comparison.

The project was aimed at building a platform providing users with personalized for-rent and for-sale listings. With an appealing interface, easy navigation, and numerous advanced filters for accurate property searching, the marketplace ensures smooth user experience and tailored property matches.

The solution comprises exceptional analytical capabilities:

• When registering, the user specifies essential personal preferences by gradually wading through suggested categories
• After registration, the algorithm evaluates every placed property to provide personalized listings indicating exact matching percentage

The platform is developed to streamline property selection by considering the customer’s behavioral patterns. By analyzing the user’s behavioral patterns in combination with general market attributes (stock dynamics, interest rates, unemployment rates, and else), the marketplace provides precise property recommendations.

The patterns are defined by the following factors:

• Liked and disliked properties
• Search attributes
• Personal information (age, gender, occupation, education, marital status)
• Personal preferences
  • Property location, size, price, house/room/lease type, renovation history
  • Landscape view
  • Chosen points-of-interest (schools, parks, city center, sports facilities)
  • Neighborhood details (average age, average income)
  • Public transportation
  • Weather conditions (local mean annual temperature)

As the project requirements constantly changed during development, we went with the Scrum methodology. This way, we could incrementally deliver business value through continuous feedback loops.

Our team has covered:

• Software development
  • Cloud architecture
  • Application architecture (frontend, backend, data storage)
  • Data architecture
  • MVP composition
• ML training and implementation

Solution architecture

At this project stage, our team took over:

1. SPA hosting

The built single-page application, based on React Native, was hosted by utilizing Amazon S3 and CloudFront. This approach ensures availability and better user experience, as all static files are downloaded at notably higher speed.

2. Dynamic user requests, microservices

• The dynamic user requests are implemented by using Amazon ECS
• Node.js-based microservices are managed by using Amazon Fargate container groups in combination with Application Load Balancers

This approach ensures scalability, network redundancy, and availability.

3. User authentication

User authentication is based on the JWT standard ensuring secure information transmission.

4. Data storage

Data storage is implemented by using Amazon RDS PostgreSQL database to manage home data, search filters, and other relevant information.

5. Advanced search and analytics

AWS OpenSearch (Elasticsearch cluster) is implemented to store and index user activities, home data, and logs, which are then used to train machine learning math algorithms, which calculate home scores and ratings.

6. Data warehousing

ETL process is implemented to extract relevant data from available external resources and load these into Amazon RDS and Elasticsearch.

ML implementation

At this project stage, our team took over:

1. Solution design

Which included:

• Requirement review
• Thorough analysis
• Metric definition
• Benchmark preparation
• Architecture design
• Development specification

2. Data analysis

During this development phase, we handled:

• Data analysis (source, volume, format, structure)
• Visual analysis (distributions, counts, PCA, tSNE)
• Data assessment
• Data augmentation

3. Data preparation

This stage comprises thorough data cleaning, data transformation and augmentation.

4. Model development

This stage includes suitable technique selection and application, as well as optimal parameter calibration.

5. Result evaluation

6. Automated learning and monitoring

The challenges our team has faced:

1. Ingesting data from various third-party resources

The obstacles associated with data ingestion were caused by different data formats and legacy source APIs. This challenge could be successfully handled by investigating various possibilities of smooth data extraction.

2. Fitting suitable scoring system

The platform includes a scoring system for categorization and ranking, which evaluates:

• Property description (content completeness, photo attachments)
• Consumer interest scores
• Community activity scores

It was quite complicated to develop a well-fitting scoring system to the great variety of the analyzed datasets. To handle this challenge, we created a multi-scoring system that considers the types and quality of datasets automatically ranking this data.

We delivered a platform that generates relevant recommendations to facilitate property search and matching. The marketplace, enhanced by machine learning, is an appealing solution streamlining greater customer reach and higher market competitiveness.