How AI Plan Review Reads Construction PDFs

PDFs are the lingua franca of construction. They're what gets permitted, bid, and built. While BIM models are valuable for design coordination, the actual construction documents—the PDFs that contractors receive, that cities approve, that trades build from—are what matter most for catching issues.

Models and drawings often drift apart. Details get added in CAD that never make it back to the model. Dimensions are adjusted. Notes are revised. The PDF is the single source of truth for what the field builds.

• Vector paths: Lines, polylines, arcs, and hatches defined mathematically
• Raster tiles: Scanned elements, images, and photos
• Text layers: Fonts, dimensions, notes, and annotations
• Metadata: Page structure, layers, and document properties

We extract everything the PDF contains—combining vector geometry with raster analysis and text understanding. This lets us understand the document as a whole, not just as pixels on a page.

Vector geometry refers to lines, polylines, and curves defined mathematically—not as pixels. When you draw a line in AutoCAD or Revit, it's stored as precise start and end coordinates, not a raster image.

This matters because vector data gives us:

• Exact coordinates, not approximations
• Scale-independent precision
• Geometric relationships (parallel, perpendicular, intersecting)
• Containment and boundary detection

Image-based AI (treating drawings as pictures) can recognize patterns, but it can't measure precisely. Vector extraction lets us:

• Detect dimension values directly from the geometry
• Identify walls, doors, and windows from linework patterns
• Calculate clearance zones with precision
• Verify grid line alignment across sheets

Construction drawings have dense, small text—often rotated, scaled, or overlapping. Standard OCR struggles with:

• Dimension strings with fractions and special characters
• Symbol labels and keynotes at various angles
• Handwritten annotations (in some cases)
• Text overlaid on hatching or backgrounds

We don't just extract text—we understand what type of text it is and how it relates to nearby geometry:

• Dimensions: Associated with specific walls, openings, or clearances
• Notes: Linked to keynotes, specification sections, or details
• Labels: Connected to rooms, equipment, or systems
• Titles: Identifying sheet names, drawing types, and scales

No two firms use identical symbols. A door might be shown ten different ways depending on the CAD standard, the drafter's preference, or the project type. The same electrical receptacle could be a circle, square, or custom graphic.

We detect symbols contextually, not by matching against a fixed library:

• Annotation context: "6-15R" near a receptacle shape = 15A receptacle
• Label interpretation: "GFCI" near outlet = GFCI protection required
• Pattern recognition: Door swings, equipment symbols, plumbing fixtures

Construction documents are interconnected. A floor plan references details, sections reference elevations, schedules reference plans. These references are how information flows—and where coordination breaks down.

• Detail callouts (e.g., 'See A5.2/3')
• Section marks pointing to section sheets
• Elevation references
• Schedule references to specific items

We parse callout references from text and symbols, then map them to actual sheets in the set. This creates a navigation graph that lets us:

• Detect broken references (callout to non-existent detail)
• Compare referenced details to plan context
• Check dimension consistency across views
• Validate schedule entries match plan counts

Given a project location and building type, we determine the applicable codes:

• Location: Determines jurisdiction (IBC, CBC, NYC Building Code, etc.)
• Building type: Determines occupancy classification and requirements
• Detected elements: Trigger specific code sections (stairs, ramps, egress, electrical)

We maintain a library of 120+ building codes including IBC, IRC, IFC, IPC, IMC, NEC, CBC, FBC, CALGreen, NFPA 13/72/101, ASHRAE 90.1, ADA, A117.1, and state-specific codes. When we detect a relevant condition (e.g., a stair, an electrical panel, an accessible restroom), we retrieve the applicable requirements.

Generating a drawing with AI is extremely hard. If it's 80% correct, it's still unusable—it must be redesigned to fully meet code, coordination, and constructability requirements.

Checking a drawing is a perfect AI use case. If AI catches 4 out of 5 real issues, that's immediate, tangible value: fewer RFIs, fewer revisions, less rework, less risk.

You're not asking AI to replace engineers or designers. You're asking it to do what machines are great at: reviewing large amounts of information, spotting inconsistencies, and flagging risks humans might miss.

In probabilistic workflows, 80% accuracy is a superpower. In deterministic ones, it's a failure. This distinction matters—and it's where AI delivers ROI today, not someday.

We'd rather miss an edge case than flood you with noise. Every finding requires multiple signals before flagging, and we maintain confidence scoring on every finding.

Every finding is designed for human review:

• Specific sheet references and locations
• Visual evidence (cropped views of the issue)
• Code citations where applicable
• Ability to mark false positives for learning

We learn from production data—common false positive patterns, edge cases that need human judgment, and new drawing conventions. The system improves over time.

Do you use computer vision, LLMs, or both?

Is this rule-based or learned?

How do you handle CAD exports vs scanned drawings?

What if my drawings use non-standard symbols?

How do you stay current with code changes?

Can this work with BIM models directly?

Is 3D/BIM support coming? What's on the roadmap?