Machine Learning Infrastructure for Commercial Real Estate Insights Platform
By Rinat Gareev, ML Solutions Architect – Provectus
By Benjamin Gardiner, Partner Solutions Architect – AWS
By Harold Li, Sr. Director of Data Science – VTS
VTS is a commercial real estate leasing and asset management software and data company that delivers an easy-to-use workflow and insights platform that empowers commercial real estate professionals to work smarter, not harder. Over 12 billion square feet of commercial real estate is managed on the VTS platform.
Founded by real estate professionals who have experienced the challenges facing today’s landlords and brokers, VTS’s mission is to be commercial real estate’s decision-making platform, where real-time insights come to life.
To do so, VTS wanted to productionize machine learning (ML) models while gaining the ability to build new models iteratively using Amazon Web Services (AWS). They wanted to accelerate the time to market for ML applications, and reduce human errors and the effort from their data science team.
Provectus, an AWS Premier Tier Services Partner with the Machine Learning Competency, looked into how ML models were prototyped and evaluated at VTS and delivered a template-based solution enabling their data scientists to more easily create Amazon SageMaker jobs, pipelines, endpoints, and other AWS resources.
The resulting coherent set of templates, with usage cookbook and extension guidelines, was applied successfully on a machine learning model that predicted leasing outcomes.
By implementing the templates for VTS, Provectus helped significantly in reducing the amount of manual activities related to bringing their proof of concept (PoC) models to production in a cost-efficient and scalable manner, with architectural best practices. This created a strong foundation for the VTS team, enabling them to develop and deliver future ML applications faster and at scale.
One of the machine learning models the VTS data science team prototyped was around predicting leasing outcomes. By surfacing these predictions with customers, landlord reps and asset managers could make more informed decisions around their leasing and investment strategies.
However, VTS faced hurdles when integrating the predictive model into the core user experience. The data scientists were capable of delivering the model in ad hoc environments (such as Jupyter Notebooks), but found it challenging to deploy the model in production using the existing infrastructure of the VTS platform.
In short, VTS had superb data scientists but lacked the necessary AWS and MLOps expertise to finish the job. As a result, VTS joined forces with Provectus to deliver on these objectives in an expedited manner.
Template-Based Model Productionization
To reduce these complexities, Provectus suggested data scientists could use VTS-specific templates and software developer kit (SDK) to create jobs, pipelines, and endpoints in Amazon SageMaker and other AWS services.
The templates and SDK would be designed, built, and delivered by Provectus, and the VTS data science and ML engineering team would receive extensive training and guidelines on how to use the templates and SDK.
Figure 1 – Stages of the model productionization.
The first set of templates provide the foundation for refactoring data ingestion, data preprocessing, feature engineering, and model training scripts into Amazon SageMaker jobs.
The delivered SDK, named VTS ML Commons SDK, extends Amazon SageMaker SDK with components tailored for VTS environments. It reduces the amount of boilerplate code and hides configuration complexity to be provided by data scientists for using Amazon SageMaker capabilities. It also encourages clear separation of data and ML processing logic from the underlying infrastructure.
The second set of templates uses this foundation to simplify connecting different workflow steps into Amazon SageMaker pipelines for training, batch-mode, and real-time model serving, as well as integration with other AWS services, like Amazon EventBridge or AWS Lambda.
Provectus also prepared the required documentation and held several knowledge transfer sessions with VTS ML engineers, to demonstrate the speed and efficiency of using the templates instead of the manual process.
Facilitating the Delivery of ML Applications
With the help of Provectus, VTS received the tools needed to make great strides in their machine learning work—from deploying existing models to developing and operationalizing new models.
The leasing outcomes model was deployed in production in an agile framework, with close collaboration between the Provectus and VTS teams. The training pipeline of this model is shown in Figure 2 below.
Figure 2 – Training pipeline of the leasing outcomes model.
The main result of the training pipeline is a model package in SageMaker model registry. This package is used for both model serving modes: real-time and batch. Training and batch mode inference pipelines are implemented as Amazon SageMaker Pipelines.
For real-time inference, SageMaker hosting services are used. All these resources are connected by rules defined in Amazon EventBridge that trigger AWS Lambda functions when the specific event occurs, as shown in Figure 3.
Figure 3 – Pipelines and model integration.
There are plenty of resources involved in model operationalization: code artifacts such as packages and container images, their job configurations, pipeline definitions, and rules of their executions. Their maintenance is greatly simplified by introducing continuous integration workflows, as shown in Figure 4.
Figure 4 – Continuous integration for model pipelines.
The templates and SDK for the development and operationalization of ML models were also delivered by Provectus within three months. The introduction of the templates into the ML delivery process allowed the VTS data science team to significantly reduce the amount of manual work required to productionize future ML models.
By automating part of their pipeline, the data scientists were able to prioritize high-value tasks, like model building, experimentation, and hypothesis checking, instead of spending time handling routine tasks for PoC operationalization.
The usage cookbook and extension guidelines provided by Provectus also ensured a smooth transition to a new ML delivery process, which allowed VTS leaders to see actionable results from day one.
All of the Provectus deliverables outlined in this post have enabled VTS to develop and deliver future machine learning applications faster, more efficiently, and on a larger scale. This has accelerated the company’s mission to deliver data-driven decision-making to the commercial real estate industry, bringing the sector into the quantitative era.
If you’re interested in building a machine learning infrastructure for your organization, apply for the ML Acceleration Program to get started.
Provectus helps companies in healthcare and life sciences, retail and CPG, media and entertainment, and manufacturing, achieve their objectives through AI.
Provectus – AWS Partner Spotlight
Provectus is an AWS Machine Learning Competency Partner and AI-first transformation consultancy and solutions provider helping design, architect, migrate, or build cloud-native applications on AWS.
*Already worked with Provectus? Rate the Partner
*To review an AWS Partner, you must be a customer that has worked with them directly on a project.