top of page

Versatile's AI detects activities in construction sites with TensorOps

Updated: Jul 2

Versatile, a successful startup, has already revolutionized the operations of several major construction companies worldwide. Their unique solution leverages an under-the-hook lifting accessory on cranes which collects valuable information about on-site activities from a variety of advanced sensors, and displays those insights in easily consumable reports, in an online platform. Additionally, they apply impressive data analytics and machine learning (ML) technology to analyze site activity, providing Superintendents and site teams with valuable insights on resource allocation, operational efficiency, and more. One specific application is an AI boosted scheduler that enables better utilization of the crane, by showing users planned schedule vs. automatically populated site execution. This project was developed in collaboration with TensorOps’ team of engineers and researchers.

Versatile’s Crane Scheduler

The Versatile Scheduler allows customers to plan their resource schedules, which can include crane time, subcontractors, elements, and truck deliveries, and more, in one centralized location. Then next day actuals are automatically populated in a side-by-side comparison with the scheduled activities. Customers can effortlessly review the automatically populated actuals from the previous day and compare them with their original plans to monitor and improve the schedule, based on a single source of truth. One particular resource that requires careful management is the crane, which is both highly valuable and in limited supply. Multiple subcontractors operating on the same site need to utilize the crane, and it is essential for site managers to ensure that each subcontractor receives the optimal amount of time for its usage. To assess their work effectively, they need to refer to the crane's actual usage history and identify which subcontractor was utilizing it at any given time, and see how the time could’ve been better allocated in the plan.

Determining the Subcontractor

While certain operations can be readily linked to a particular subcontractor based on the distinct actions they involve, there are other operations that cannot be associated with a subcontractor using deterministic rules. For instance, the installation of curtain walls is always carried out by the subcontractor responsible for glazing, whereas moving a tool box does not correspond to any specific subcontractor. To address this, Versatile's AI identifies the process type and then utilizes a basic model to deduce the appropriate subcontractor. There was a need to enhance the initial model with additional machine learning capabilities to cover scenarios where the process type alone is insufficient.

AI-Enabled Subcontractor Detection

Despite the difficulty of establishing a logical framework to determine the subcontractor, TensorOps and Versatile have successfully developed an algorithm that accurately estimates the most probable subcontractor. This solution relies on AI models learning the similarities between crane activities with known subcontractors and unlabeled activities. The AI system takes into account multiple signals, including GPS usage patterns, altitude, load types, and temporal features, to automatically create a profile of actions for each subcontractor. It then compares all "unclassified" crane activities with the learned patterns to make associations. For this task, tree learners are used as they can fit relatively small data with various dimensions.

Example of a learned model tree

MLOps pipelines

The learning system is implemented on Databricks and includes a pipeline triggered by the event-driven system once sufficient information from the devices has been collected and analyzed to initiate the learning process. The AI model is trained on past activity data from the site. While signals that determine load types may include unstructured data like vision, the learning model operates on processed data, receiving a tabular dataset. For instance, a computer vision algorithm may detect that the load used was concrete, but this information is utilized in a previous stage. The subcontractor association model receives the inferred load type as a feature. To evaluate the model, K-Fold validation is performed, with some crane activities designated as tests in each iteration. The overall accuracy is calculated across all known crane activities, and predictions that exceed a certain threshold of confidence score are presented to the user.

High level flow of algorithm on AWS and Databricks

Thanks to the well architected Versatile system, the algorithm leverages reduced and processed data allowing training the models and making inferences only on the aggregated analytical data and not on the original much heavier signal data. Therefore, these models return a prediction within minutes with very low running costs.


The image blow shows the effect of applying the algorithm. The effect of the prediction is of course observed only on the actual data and not on the planned (that was removed from this view). As it shows, many of the items that were considered "unknown" or "general" were associated with a contractors marked in other colors.

Results of the scheduler before and after

Overall, the work resulted in identifying correctly 85%-90% of the previously unknown task performers. Versatile’s collaboration with TensorOps yielded a combination of skilled ML researchers and consultants together with domain experts and excellent R&D capabilities at Versatile. We also found that accuracy is strongly correlated with confidence, therefore Versatile is able to only show to the user subcontractors that result from a high confidence level.


TensorOps collaborates with businesses to leverage AI, if you wish to work with us on your case don't hesitate to contact us here:

bottom of page