Turning seven years of fragmented cost data into decisions.
I worked across product definition, data cleaning, interface design, and forecasting to turn years of spreadsheets and PDFs into a usable internal decision system.
2020–26historical pricing unified
ROLE
Cost Engineering Intern · Product and data lead
WHEN
June–July 2026
WITH
Cost Engineering and Marketing teams at PertaMC
SIHMT began with an operational problem rather than a polished product brief. Seven years of material, equipment, labor, vendor, and tender data existed across Excel workbooks and PDFs, but the structure changed between files and the information was difficult to compare.
I translated those fragments into two connected product views. Cost Engineering needed searchable price references, upload review, and item-level forecasting. Marketing needed tender and vendor intelligence. The system therefore had to make imperfect historical data usable without hiding its uncertainty.
I designed one workflow for multi-file ingestion, review, normalization, search, tender analytics, and item-level forecasting, spanning both Cost Engineering and Marketing needs.
03
THE PROCESS01 02 03
How I moved through the work.
Map the real workflow
I documented how teams received files, searched for references, corrected records, and compared tenders before deciding what the interface should automate.
Build a review-first pipeline
Uploads were designed around extraction, normalization, manual confirmation, and only then saving to the master dataset. This mattered because source files were too inconsistent for blind automation.
Treat forecasting as evidence
I planned item-level baselines across ETS, Holt-Winters, ARIMA, and curve-based models, compared through R², RMSE, and MAE rather than selecting a model by appearance.
04
KEY DECISIONS
The choices that shaped the result.
- Separate the Cost Engineering and Marketing dashboards while keeping one underlying information model
- Preserve human review between OCR or spreadsheet parsing and database insertion
- Show per-item forecast quality so users can judge whether a prediction deserves trust
- Design around materials, equipment, labor, vendors, and tenders instead of forcing every record into one generic table
WHAT IT DELIVERED
Unified material, equipment, labor, vendor, and tender information
Designed review-first uploads for imperfect Excel and PDF source data
Compared forecasting approaches using R², RMSE, and MAE
MY SCOPE
Product strategy
Data cleaning
Forecasting
Frontend engineering
The hardest part was not choosing a forecasting library. It was deciding where automation should stop and professional judgment should remain visible. SIHMT taught me that useful intelligence systems are often built around careful review, provenance, and uncertainty.