A plain-language guide
The process combines visual AI, ordinary software checks, and focused review. Each part has a specific job.
Drawing intake
The pipeline accepts drawing PDFs and records the project, package, sheet number, title, revision, and page location. That source record follows the data throughout the process.
A value without its source cannot be trusted or audited later.
Visual gate
Qwen 3.5 looks at the sheet, distinguishes schedules from diagrams and notes, identifies the table type, and returns candidate records in a defined format.
Recent testing suggests direct visual understanding is a stronger first gate than relying only on handcrafted line and whitespace detection.
Deterministic checks
Software checks data types, required fields, units, allowed values, duplicate rows, cross-references, and impossible combinations. For example, a door number remains text because “101A” is valid.
AI should not spend time solving checks that ordinary rules can perform consistently.
Focused review
Records that fail a check—or contain unclear cells—can receive a second model review. A person only needs to inspect the smaller set that remains unresolved.
The goal is not blind automation. It is faster review with visible evidence.
Data twin
Doors, hardware, finishes, equipment, and electrical schedules become queryable records. Every record retains its relationship to the source drawing and can later connect to BIM elements without overloading the model.
The extraction is only useful when people can apply the result to estimating, procurement, coordination, and field work.
Design rules
Every value keeps its source sheet and location.
Blank or illegible fields stay blank and enter review.
Detection, rows, fields, normalization, and review are scored separately.
The schema and checks survive when the underlying AI model changes.