Granite Construction · Safety Intern · Pacific Northwest · Summer 2026
The information was already in the reports. Nobody was reading across them.
I spent a summer as a safety intern at Granite Construction, one of the largest civil contractors in the U.S. My job was to help a regional safety team see what the field was already telling them: I went to jobsites, read incident narratives one at a time, built classification workflows with a human sign-off on every row, and turned the results into decisions leadership could act on.
Four projects, one through-line: get the right information to the right person before something goes wrong.
STCKY Luck vs. Success
58 times this year someone could have died. The reporting counted two of them.
Leadership's headline safety ratio was built from 8 tagged events. I read every report in the region instead, with a strict decision tree and a person approving every call, and rebuilt the number from evidence.
Read the case studyWork Zone Intrusion Classifier
Turning 143 messy near-miss narratives into a spec a development partner could build against.
"Car went around flagger." "Truck drove through job site." Real intrusions were buried in a general log with no shared definition. I wrote one, built the pipeline, and reviewed every row.
Read the case studyJob Site Management Hub
One place to find every active site, who to call, and air-quality alerts by text.
Site addresses, meeting points, contacts and PPE rules lived in emails and people's heads. I built a Power Apps hub fed by three automated flows, then handed it to the enterprise team to ship.
Read the case studyLight-Vehicle Incidents & Field Work
Moving a supervisor conversation from driver blame to exposure management.
Every 2026 light-vehicle incident in the region, presented at a supervisor SHEQ meeting, plus the jobsite inspections and observations that kept every data project honest.
Read the case studyHow I worked
Research first. Then a definition, a machine, and a person who signs off.
Go to the site
Every data project started on a jobsite. Inspections, observations and a SHEQ assessment showed me what a "near miss" actually looks like before I tried to classify one.
Write the definition down
Most of the disagreement in safety data comes from undefined words. I wrote one explicit definition per project and applied it the same way to every record.
Let a person decide
The model reads and explains; a reviewer approves, excludes or escalates. Nothing reached leadership without a human decision behind it and a source anyone could check.
Tools I used
- Power Automate
- Copilot Studio
- AI Builder
- Power Apps
- Dataverse
- SharePoint
- Excel
- HCSS incident data
- PowerPoint
Why this belongs in a UX research portfolio
None of these projects started with a tool. They started with people who needed a clearer picture: a safety manager defending a number, a supervisor deciding where trucks park, a crew member trying to find the meeting point on a site they'd never visited. The method was the same one I use in UX research: observe, define, gather evidence, and present it so someone can act.