How IMRIS turned 40,325 device history files into verified maintenance data with tools built in Claude Code
A surgical imaging manufacturer needed five data points from several hundred signed maintenance reports buried in its device history records. After Shannon Advisory set Claude up inside the company's controlled Microsoft environment, a manufacturing engineer in the pilot group built two reusable tools with Claude Code that found the reports, extracted the data, and traced every value to its source.
Bars to scale; the lighter segment shows the range IMRIS estimated. IMRIS's figures, from its July 2026 write-up of the results.
- 316 preventive maintenance reports identified among 40,325 files in 5,664 folders of device history records.
- 1,280 data points extracted, five fields per report, each traceable to its source document through a live tracker and log sheet.
- Extracted values matched the signed PDF reports when IMRIS checked them against the source documents.
- About 53 to 79 hours of manual extraction replaced, on IMRIS's estimate.
- Two reusable tools built with Claude Code; a future extraction from the same records needs only new keywords and fields.
The starting point
IMRIS makes surgical imaging systems installed in hospitals, from its base in Chaska, Minnesota. It performs recurring preventive maintenance on every installed system and documents each visit in a signed PDF service report of about 13 pages.
For system lifecycle and utilization analysis, engineering needed five data points from each report: the hospital or site, the maintenance period, the overall pass or fail result, and the linear and rotator cycle counts. The reports sit inside IMRIS's device history records: 40,325 files across 5,664 folders, some text, some scanned, some handwritten. The first problem was finding which documents held the data at all.
An earlier approach had taken about eight Python scripts to locate and copy the right PDFs and extract each field, with separate versions for text-based and scanned documents, and it needed human verification at every step.
Claude inside a controlled environment
IMRIS engaged Shannon Advisory for a four-week implementation that began on June 29, 2026, with three days on site in Chaska followed by structured virtual support. IMRIS runs a tightly controlled Microsoft environment with real constraints on what any new tool may touch, so Shannon configured Claude inside the company's existing identity and tenant, working directly with its systems administrator, with access governance in place from day one.
The on-site days went to training the teams on their own work. A pilot group then took Claude into daily use, with office hours for support. The development team works with Claude in their IDE, and a revenue forecast skill built live with the finance team populates monthly actuals into the master forecast.
Two tools built with Claude Code
Lavanith Togaru, a manufacturing engineer in the pilot group, took the maintenance-report problem to Claude Code. His goal was a versatile, robust tool whose output could be verified as easily as possible.
From a couple of sample documents and a description of what he needed, Claude Code helped him build two Python tools. The first searches thousands of folders for keywords and copies the matching documents, whether they contain text, scanned images or handwriting. The second extracts the fields, shows the values live as it works, and writes them to an Excel workbook with a log sheet.
It took a couple of iterations, mostly run by Claude Code in the background while Togaru worked on other tasks. His part was checking each iteration and saying what needed tailoring to his needs.
Checked against the source
For the five data points needed from each document, the tools identified 316 PM reports and extracted 1,280 data points; some documents were samples with no applicable data. IMRIS checked the extracted data against the signed PDFs and the values matched. The live tracker and the log sheet made every value traceable to the document it came from.
IMRIS put the manual equivalent at 53 to 79 hours of extraction work. Because the tools are not tailored to maintenance reports, the next extraction of any documented data needs only new keywords and fields.
What IMRIS's rollout shows
- Set up the environment before the pilot.Claude went in inside IMRIS's own identity and tenant, with access governance from day one, so the pilot group could work with real device records.
- Ask for a tool, not a one-off answer.Togaru described the tool he wanted rather than the single extraction. The result works on any documented data, not only this set of reports.
- Build for verification.Showing values live and logging each one against its source is what let IMRIS check 1,280 values against the signed PDFs and use the data for engineering analysis.
- Let Claude Code do the iterating.Most iterations ran in the background. The engineer's time went to checking each round, not writing the scripts.
What's next
The same two tools can be pointed at other data in IMRIS's device history records with new keywords and fields, without new code. As more IMRIS teams put Claude to work, Shannon expects to help carry the same approach into their workflows.
In their words
"My goal was never just to extract the data. I wanted to build a robust, versatile tool that can give me a desired output I could verify as easily as possible, and Claude Code let me build that just by describing it."Lavanith Togaru, Manufacturing Engineer II, IMRIS
About this case study. Published with IMRIS's approval, September 2026. The account of how the tools were built, and the quote, are Lavanith Togaru's own, from his written description of the work in September 2026. Figures are IMRIS's own, from its July 2026 write-up of the results, checked against the source PDFs; the hours figure is IMRIS's estimate.