Digital Technology / Data Analysis Intern, June to December 2026
Seven projects, in the order I undertook them. Each one records how far it got, because a system that was built and a system that was used are different claims. Internal names, data and colleagues are left out. Screenshots show real interfaces running on sample data, and the support assistant is shown as a schematic.
What I owned was the problem framing, the experiment design, the validation protocol, and every keep or reject decision.
From my internship notes, on working with AI coding assistants
Tested whether machine logs could time tool changes better than a piece counter.
MethodValidation used cycle-grouped splits over 5 and 20 seeds, a time-based split, walk-forward folds and leave-one-machine-out.
Interface decisionEach tool receives a risk band (low, watch, medium, high), and every warning is advisory. The system never changes or extends a tool on its own.
Audited a vendor forecast, then built and backtested a simpler one.
MethodThe audit came first. Five of the six existing models had been trained on a quantity other than shipped units, and their test sets reused validation data, so there was no fair benchmark to beat.
Interface decisionEach result was delivered as one self-contained HTML page with the figures, charts and caveats together, so the planners could open it in a browser without running any code.
Moved daily order workbooks from SharePoint into a Fabric lakehouse.
MethodThe first design was an 18-step Power Automate flow that wrote to OneLake through its API. I resolved eight faults before reaching one that would not move.
Answers after-sales questions in plain language and cites its source.
MethodRetrieval combines BM25 keyword matching with Qdrant vector search, merged by reciprocal rank fusion. Exact model codes receive extra weight.
Interface decisionTyping / lists the questions the assistant can answer, so its scope is visible rather than guessed. I also cut the first-run guide from 329 words to five short slides.
Drafts a hazard analysis from two photographs and a one-line note.
MethodTwo vision calls run in sequence. The first builds the cause tree and the second converts it into the 5M1E grid. The schema marks every category as required, so none can be skipped.
Interface decisionEvery field has a character counter sized to the cell it lands in, so text that fits on screen also fits in the workbook.
Reads tolerance values from scanned drawings for a person to confirm.
MethodThe model reports only what it sees and where. Plain Python assigns each value to its point, so a top and bottom swap cannot happen.
Interface decisionA user asked why "Confirmed" showed 44 when "All" showed 40. That question led me to redesign the tabs.
Rebuilds a planning workbook's figures as a web application.
MethodIt rebuilds all 76,028 monthly cells from their order lines and matches the workbooks on 37 of 37 review groups.
Interface decisionThresholds set by the planner are labelled apart from thresholds set by the app, so it is always clear whose number is whose.