Transforming Safety Management at Suffolk with Predictive Analytics

According to the Bureau of Labor Statistics, over 150,000 construction site injuries occur every year. To mitigate incidents before they occur, Suffolk’s safety and data teams joined forces to develop a safety model that integrates data on staffing, trade partners, observations, incident history, project schedule, and core project details to proactively assess project risk.

“Over the years, we have made significant progress in equipping every employee with real-time predictive insights,” said Aleksey Chuprov, VP of Data Analytics and AI at Suffolk. “Those insights power optimal decision-making and best-in-class project outcomes across our seamlessly connected, analytics-driven platform.”

Using Posit Team, the data science team at Suffolk standardized their model workflow and communicate safety scores to stakeholders in two separate ways:

• Safety Dashboard to provide key insights (percentage of high risk projects, any projects with incidents, how predictions compare to historical predictions over time, etc.) at both the project and portfolio level to help project teams see risk in a few different ways
• Weekly Portfolio Risk Email with a concise list of projects flagged as high risk, plus the top features driving each score

When Blake came onboard, much of the data science work had grown organically and there was no cohesive enterprise-grade infrastructure to develop and deploy solutions for the team. Their prior workflow was dependent on local Alteryx scripts, and they ingested a fair amount of one-off data sources (e.g., data that had been extracted and manicured in Excel).

As a result there were bottlenecks on the execution of the model, as well as questions on the lineage and reproducibility of the data. Often with the data being ingested from Excel, it was not able to be replicated from its original ‘sources’.

Suffolk needed a solution for seamlessly communicating their analyses and models created in Python and R with business teams. While existing dashboards could support static data, they needed a reproducible solution that would allow site managers and project executives to interact with deployed models to make decisions backed by data.

Blake Abbenante estimates that by moving to Posit, the team has “reduced deliverability time by ~20% simply by not having to search through SharePoint for Excel files.”

Stakeholders may place more trust in their subjective experience than what data is telling them.

This isn’t necessarily pushback, but they know better as there’s nuance in the actual management of a project that is not always captured in the available data.

Generally, before introducing a new project, identify the key stakeholders (site manager, superintendents, Project Executives) who have more intimate knowledge of projects themselves to get their insight and feedback. They can then be a part of the initial rollout.

In partnering with stakeholders, be upfront that there’s probably data that is currently not represented. To continuously enhance our models, we need to gather more nuanced data. This requires collaboration with stakeholders (those on the job sites) to determine how to collect the necessary level of detail. Then by incorporating this additional data into future iterations, we can consistently improve the model's fidelity.

With the partnership of Databricks and Posit, Suffolk leverages row-level permissions in their interactive applications. This ensures business stakeholders can use their own credentials to access a personalized view with only the data they should see through a single, shared app.

For a more technical detailed walkthrough of how this is made possible with Posit Team + Databricks, you can view Max Patterson’s Workflow Demo below:

For the safety model, the way the data was presented was very important. The initial framing of the data as “having a higher likelihood of incident” made it something project teams wanted to avoid. This can have an impact on how people collect the data or tune out the results, just by how the solution is framed. It’s crucial to clearly communicate to stakeholders what you’re doing and how it can help make them more efficient.

The future

Suffolk aims to be the future leader of AI in the built world and it’s a company-wide initiative to figure out how they can best leverage AI. Some of the initial use cases so far have been chat agents to retrieve specific information about particular domains or extracting data out of PDFs, but future use cases of AI are to augment and speed along design of buildings themselves and in creating entire schedules for project planners. If it may take a planner 2 weeks now to do the scheduling planning, is there a way to get them 80% of the way in a matter of hours or minutes.

For data scientists interested in taking the first step into AI, Blake recommends checking out the ellmer and chatlas packages from Posit. He adds, “It’s really easy then to integrate it to any other kind of work that you’ve done or if you happen to be using any of the Posit suite of tools, so building it into a Shiny app or whatever else. I think that is probably the easiest way to get started.

Director of Data Science and Analytics at Suffolk