Versatile's AI detects activities in construction sites with TensorOps

You are absolutely right. In the effort to make it "punchy," I over-compressed the technical nuances and the specific operational context that makes the Versatile story compelling. The "curtain wall vs. toolbox" example and the specific architectural decisions (raw vs. tabular data) are crucial details.

Here is a comprehensive, detailed revision that retains the professional tone but restores the depth, specific examples, and technical richness of the original text.

Case Study: Solving the "Black Box" of Construction Logistics with AI

Versatile has revolutionized the global construction industry by turning the job site’s most critical asset—the crane—into a smart IoT device. Their solution leverages a proprietary under-the-hook lifting accessory packed with advanced sensors to collect real-time data on site activities.

However, collecting data is only half the battle. To truly empower Superintendents and site teams, Versatile partnered with TensorOps to develop an AI-driven Crane Scheduler. This project bridged the gap between raw sensor data and actionable business intelligence, solving a complex attribution challenge that had previously left site managers in the dark.

The Challenge: The "Unknown" Variable in Site Scheduling

The crane is the heartbeat of any vertical construction site. It is a limited, high-value resource that must be shared by multiple subcontractors (glaziers, concrete teams, steelworkers).

The Versatile Scheduler was designed to be the "Source of Truth," allowing customers to view their Planned schedule side-by-side with Actual execution. While the platform could track what the crane was doing, it struggled to identify who was doing it.

• The Deterministic Limit: Some actions follow strict rules (e.g., installing a curtain wall is always done by the glazing subcontractor).
• The Ambiguity Problem: Many crane movements—like moving a toolbox, lifting generic materials, or repositioning equipment—cannot be assigned to a specific team using simple logic.

Without AI, these ambiguous lifts appeared as "Unknown" or "General" in reports. This gap prevented managers from accurately validating subcontractor invoices or optimizing future crane allocation.

The Solution: AI-Enabled Subcontractor Detection

To fill these gaps, TensorOps and Versatile developed a sophisticated Machine Learning algorithm capable of "deducing" the subcontractor behind every lift.

Instead of relying on rigid rules, the AI builds a behavioral profile for each team. The model analyzes a complex matrix of signals to find the "fingerprint" of a subcontractor’s activity, including:

• GPS & Spatial Patterns: Where the load is picked up and dropped off.
• Altitude Data: The specific floors or heights where a subcontractor operates.
• Inferred Load Types: utilizing data from computer vision algorithms (e.g., distinguishing concrete from steel) as a feature input.
• Temporal Features: The time of day and duration of specific lift types.

By comparing these signals against known patterns, the system can accurately classify previously "Unknown" activities.

Versatile's Scheduler demo page

Technical Deep Dive: A Cost-Efficient MLOps Architecture

The engineering challenge was not just accuracy, but efficiency. The solution was built on AWS and Databricks, employing a smart pipeline designed to minimize compute costs while maximizing speed.

1. The "Tree Learner" Approach TensorOps utilized Tree Learner models for this classification task. This choice was strategic: Tree Learners are exceptionally good at fitting datasets with relatively small sample sizes but high-dimensional complexity, making them perfect for the varied and noisy environment of a construction site.

2. Smart Data Processing A key architectural win was the decision to decouple the heavy signal processing from the inference engine.

• The system uses an event-driven pipeline triggered only when sufficient device data is collected.
• Crucially, the ML model does not train on heavy, unstructured raw data (like video feeds). Instead, it ingests a processed tabular dataset. For example, while a computer vision algorithm does the heavy lifting to detect that a load is "concrete," the Subcontractor Model simply receives "concrete" as a data feature.
• Impact: This allows the model to return predictions in minutes with very low operational running costs.

3. Rigorous Validation To ensure reliability, the model undergoes K-Fold validation, where specific crane activities are held back as test cases in iterative cycles. The system also calculates a Confidence Score for every prediction. To maintain trust with the user, Versatile only displays subcontractor attributions that exceed a strict confidence threshold.

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.

Results & Business Impact

The collaboration between TensorOps’ ML researchers and Versatile’s domain experts transformed the Crane Scheduler from a passive tracker into an intelligent analyst.

• 90% Identification Rate: The algorithm now correctly identifies the subcontractor for 85% to 90% of previously "unknown" tasks.
• Visual Clarity: In the Scheduler interface, the "Unknown" grey blocks were replaced with color-coded subcontractor assignments, giving Superintendents an immediate, accurate view of resource utilization.
• Actionable Insights: By correlating accuracy with high confidence scores, the system ensures that site managers are making decisions based on reliable data, allowing for better dispute resolution and schedule optimization.

Results of the scheduler before and after

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