Problem
Many businesses hold important information in PDFs, technical reports, tables, diagrams and visual plans. That information can be hard to reuse because it is locked in documents made for people, not for databases or digital workflows.
In this project, the challenge was to extract structured information from complex Gantt-style planning diagrams in an industrial context. Users previously had to compare documents manually to spot changes in activities, dates, colors, shapes, status indicators and rows.
Solution
I developed a modular Python pipeline that combined PDF processing, layout analysis, computer vision, color detection, shape recognition and rule-based interpretation. The system turned visual plans into structured data that could be exported to CSV, Excel, JSON or SQL-ready formats.
OCR was tested but was not stable enough for this document type. The final solution was therefore built around more controllable methods: deterministic PDF extraction, image processing, geometric analysis and clear validation rules.
Result
The system was evaluated with manual comparison, train/test-style validation and strict correctness checks on roughly 150 pages. Reported accuracy was above 95% under a strict definition where the entire row had to be correct to count.
The project shows how unstructured, visual and domain-specific information can be turned into practical, database-ready data. For a business, this can reduce manual work, speed up comparisons and enable analysis across many documents.