All case studies
Case study 05AI SystemsAutomationMachine Learning

Engineering Drawing Review Agent

42 deterministic checks across DXF, vector PDF, and scanned drawings, grounded in ASME Y14.5.

Role
Senior engineer · end-to-end build
Core stack
LangGraphTesseract OCRYOLOv8RAGPython
42
Deterministic checks automated per drawing
3
Input formats: DXF, vector PDF, and scans
2
Report formats generated: Word and Excel

Overview

A LangGraph agent that reviews engineering drawings automatically, combining OCR, object detection, and RAG over technical standards, and produces review documents for engineers.

Drawings arrive in mixed formats, from clean CAD exports to scans. Reviews must be consistent and explainable, reference the governing standards, and cross-check related records such as revisions, BOMs, and purchase orders.

Product walkthrough

The flow, screen by screen

Drawing review screen: an assembly drawing with numbered finding markers beside a findings panel showing severity, confidence, and the standard each finding is based on
01 / 05

Drawing review. Each drawing opens with its findings pinned to the sheet. Every finding shows its severity, a confidence score, and the standard it is based on, and engineers accept, reject, or annotate it in place.

ArchitectureCorePlatformAsync / control
asyncDXF filesVector PDFsScanned drawingsGeometry parsingDXF · PDF vectorsOCR + detectionTesseract · YOLOv8Review agentLangGraph42 checksdeterministic rulesStandards RAGspecs · ASME Y14.5Cross-checksrevisions · BOM/PO · STEPReportsWord · ExcelCelery workersRedis · retries
Inputs are normalized and analyzed, then the agent runs deterministic checks, standards retrieval, and cross-checks before generating reports.

Approach

  1. 01Built a LangGraph agent that automates 42 deterministic checks across DXF, vector PDF, and scanned drawings.
  2. 02Extracted text and symbols with Tesseract OCR and YOLOv8, and used RAG over technical specifications and ASME Y14.5 to ground findings.
  3. 03Integrated revision comparison, BOM/PO and STEP cross-checks into the review workflow.
  4. 04Generated Word and Excel review documents automatically for engineering teams.
  5. 05Ran ingestion, analysis, and report generation on Celery, Redis, and Docker with task routing, retries, and failure recovery.

Key decisions

Deterministic checks first, LLM second

Rule-based checks give repeatable results; retrieval over the standards explains why a finding matters.

One pipeline for every input format

Normalizing DXF, vector PDF, and scans early lets every check run the same way regardless of source.

Queue-based processing

Celery workers with retries keep long-running drawing analysis resilient and scalable.

Outcomes

  • Findings reference specifications and ASME Y14.5 for traceability.
  • Workflows tested for failure recovery before hand-off to client teams.
  • LangGraph
  • Tesseract OCR
  • YOLOv8
  • RAG
  • Python
  • Celery
  • Redis
  • Docker

Working on something similar?

I work with clients worldwide, across any time zone. Tell me what you're building.

Send a message Email meMore case studies