AI-Powered Construction Takeoff Automation for a Leading Construction Technology Platform Company
AI-Powered Construction Plan Analysis
Client
Construction Technology Platform
Industry
Construction Technology
AI-Powered Construction Software
Service
Construction AI Automation
Computer Vision & Document Intelligence
Use Case
Automated Construction Takeoff
Construction Plan-Set Analysis
Technology
YOLOv11
GATv2 / FloorPlanGAT_Pro
Goal
Address the client’s requirement for an automated construction takeoff workflow capable of converting complex plan-set PDFs into structured, measurement-ready data. Vedya Labs was tasked with reducing the manual effort involved in identifying relevant sheets, extracting sheet information and drawing scales, detecting construction elements, and preparing takeoff data for integration into the client’s platform.
Problem Statement
The client’s construction workflow relied heavily on manual analysis of multi-discipline plan-set PDFs. Estimators needed to inspect individual sheets, identify architectural floor plans, interpret title blocks, determine drawing scales, locate construction elements, and translate drawing information into usable quantity records.
For the client, automating this workflow presented several technical challenges. Construction plan sets contained different disciplines, inconsistent sheet formats, varying scale notations, large high-resolution drawings, and thin interconnected elements such as walls, doors, and windows. In addition, the resulting AI output needed to integrate directly with the client’s backend platform rather than remain as standalone visual detections.
Solution Highlights
Vedya Labs addressed the client’s requirements by implementing a dedicated AI-powered takeoff pipeline tailored to its construction plan-processing workflow. The implementation combined PDF document intelligence, computer vision, graph-based geometry reasoning, scale extraction, structured data generation, and backend orchestration.
Construction Plan-Set Intelligence
To automate the client’s initial plan-set analysis, Vedya Labs implemented PyMuPDF-based document intelligence to extract sheet numbers, disciplines, sheet names, sheet types, scale information, page geometry, and takeoff eligibility directly from the submitted PDFs.
Automated Sheet Identification
The client required the system to identify only the sheets relevant to its takeoff workflow before applying computationally intensive AI processing. Vedya Labs implemented sheet classification and eligibility filtering to identify architectural floor-plan sheets and exclude non-relevant pages.
The implementation achieved 100% accuracy for sheet number and discipline extraction across 188 benchmarked sheets.
Scale Extraction for Measurement
Since accurate quantities depend on the correct drawing scale, Vedya Labs implemented scale extraction to identify architectural inch-to-foot notation and NTS variants from the client’s plan sets. The extracted scale was converted into points-per-foot to support downstream measurement conversion.
The implementation achieved 89.36% scale-text accuracy across the 188-sheet benchmark, with 100% targeted accuracy defined for the Beta iteration
High-Resolution Plan Processing
The client’s architectural drawings contained detailed elements that could become difficult to detect when processing an entire sheet at once. To address this, Vedya Labs rendered eligible sheets at 300 DPI and divided them into 640 px tiles with 128 px overlap for AI inference. This allowed the system to process detailed regions of large plan sheets while maintaining coverage of small construction elements.
YOLOv11-Based Construction Element Detection
To automate the identification of construction elements within the client’s floor plans, Vedya Labs integrated YOLOv11 for candidate detection of walls, doors, windows, and other takeoff elements.
Overlapping detections generated during tiled processing were merged using IoU-based filtering to prevent duplicate measurements.
The stored YOLOv11 run achieved 0.775 precision, 0.696 recall, and 0.742 mAP@50.
Geometry-Aware Detection Verification
Visual detection alone was not sufficient for the client’s requirement for reliable construction geometry. Vedya Labs therefore incorporated a GATv2-based FloorPlanGAT_Pro model to analyze the vector geometry contained within the client’s PDF drawings.
The graph model evaluated features such as endpoints, segment length, angle, curves, and adjacency to determine whether the underlying drawing geometry supported the YOLO detections.
Multi-Model Wall Measurement
For the client’s wall measurement workflow, Vedya Labs combined YOLOv11 candidate detection with graph-based vector reasoning. This approach allowed visual detections to be checked against the actual vector structure of the drawing and established the foundation for deriving wall lengths directly from classified vector segments rather than relying only on bounding-box dimensions.
The graph model evaluated features such as endpoints, segment length, angle, curves, and adjacency to determine whether the underlying drawing geometry supported the YOLO detections.
Room-Area Extraction
Where room-area information was already available as text within the client’s construction PDFs, Vedya Labs implemented direct PDF text extraction rather than relying solely on computer vision. The system identifies area patterns such as SF and SQ FT and associates them with nearby room labels.
The graph model evaluated features such as endpoints, segment length, angle, curves, and adjacency to determine whether the underlying drawing geometry supported the YOLO detections.
Structured Takeoff Output
The client required the AI results to be directly usable within its existing platform. Vedya Labs therefore converted the processed drawing information and AI detections into structured JSON containing sheets, measurements, units, confidence values, geometry references, and diagnostic fields.
Sample outputs from the client’s available plan-set runs contained 18,450 measurements across 17 eligible sheets.
Backend & Cloud Integration
To fit the solution into the client’s existing workflow, Vedya Labs implemented a Flask-based API for takeoff requests, while Celery and Redis handled long-running processing asynchronously. Generated JSON and sheet assets were uploaded to Azure Blob Storage, and the client’s takeoff-run status was updated automatically through the platform APIs.
This allowed the AI processing workflow to operate as an integrated backend capability rather than as a standalone analysis tool.
Performance Achieved
The implementation was evaluated across 188 benchmarked sheets and achieved:
100% — Sheet Number Accuracy
100% — Discipline Accuracy
96.28% — Sheet Name Accuracy
89.36% — Scale Text Accuracy
The stored YOLOv11 run achieved:
0.775 — Precision
0.696 — Recall
0.742 — mAP@50
The implementation also generated 18,450 measurements across 17 eligible sheets in the available API outputs.
Conclusion
Vedya Labs addressed the client’s requirement for automated construction takeoff by building an AI-powered workflow specifically around its plan-set processing and backend integration needs. By combining PDF document intelligence, YOLOv11-based detection, graph-based vector verification, scale-aware processing, and structured output generation, the implementation transformed complex construction drawings into reviewable, platform-ready measurement data.
The solution provided the client with a technical foundation for reducing repetitive plan analysis while establishing a path toward more accurate geometry-based wall measurement within its existing construction workflow.