AI Drawing Extraction — Automated Quotation & Material Breakdown

Detailed Capabilities

Unified Data Model for All Document Sources

All sources — Excel, vector PDF, imperial cutlist, scan images — normalize to a unified structure. Adding new document types doesn't require rewriting calculation or quotation export logic.

SAP-Exported BOM Excel Extraction (Q2)

Reads the internal structure of files that common Excel libraries report as corrupted; recognizes columns by field name rather than position, making it resilient when column order or width changes between exports. Merges multiple document sets into a single consolidated quotation: actually ran 11 BOM files in one pass, aggregated quantities across the entire job before verification, producing a total of 374,559.89 USD with the summary section matching exactly.

Vector PDF Bill of Quantities Table Reading (Q3)

Reconstructs BOQ table rows and columns accurately based on character coordinates rather than image recognition, yielding high precision; verifies by section rather than just totals, so deviations surface at the right location. Handles Japanese terminology documents: correctly recognizes labels like 合計重量 (total weight), デッキ (deck), 手摺 (handrail) — no manual translation needed before feeding into the system.

SolidWorks Imperial Cutlist Processing (Q4)

Converts feet–inch–fraction strings to millimeters and pounds to kilograms before price calculation; all pricing calculations run on a unified metric system. Maps material standards to JIS (ASTM A36 → SS400, A500 Gr B → STKR400, S235JR+N → SS400); lines without direct JIS equivalents are flagged for reviewer decision, not auto-converted. Also calculates painting / surface treatment costs: when components lack lookup codes, the system estimates area based on m²/ton rates for section groups taken from the material sheet itself, clearly labeled as estimates.

Scan Drawing Reading Without Text Layer (Q1)

Recognizes title blocks and annotations from scanned images (e.g., UMW-550L exhaust silencer, weight 2,680 kg, material SS, heat-resistant paint 600°C), then extracts 23 structural components with materials and dimensions. Three-layer fallback, not dependent on a single provider: prioritize AI vision model → extracted results → character recognition. Can plug in Claude, OpenAI, or Gemini as preferred.

Numerical Verification Before Every File Export

Σ extracted quantities are automatically cross-checked against the reference total printed in the source document: within threshold allows export, outside threshold blocks and routes to reviewer — no seemingly-correct files are exported. Catches errors in the source documents themselves (e.g., 325 kg discrepancy in Q3 because the detail table has a handrail 手摺 line but the reference total デッキ doesn't include it).

Lookup from Customer's Own Data

Densities and standards are read directly from the material sheet in the customer's template (~1,060 lines). No external lookup tables are used, so results match how the enterprise currently calculates.

Human-in-the-loop — Clear Division of Machine vs Human Decisions

Fully automated by machine: read & extract components, convert units (imperial → metric), look up density from material sheet, calculate quantities and surface areas, cross-check Σ against reference total, fill template preserving formulas.

Machine flags, doesn't auto-decide: steel grades without JIS equivalents, lines missing quantity/dimensions, confidence below configured threshold, Σ exceeding deviation threshold, fire-resistant/stainless/fastener materials, documents exceeding template fill capacity.

Human decides, never automatic: commercial unit prices (system always leaves blank), confirm steel grades after mapping, approve flagged lines, decide when source documents conflict, approve final quotation before sending.

Each time a reviewer makes corrections, the system re-maps material standards and recalculates verification immediately, while logging who changed what — the final quotation is always traceable to its source.

Full Control of Quotation Templates

Fills into the customer's actual template (opens the correct file, preserves formulas, formatting, and layout); the original Breakdown sheet is kept intact for parallel verification. The template library has an admin page: naming, selecting default version, enabling/disabling per document type; users only select from a list when exporting, no need to upload files themselves. Changing templates doesn't require re-running extraction — template selection is separate from document reading, outputting the same dataset to a different template takes only seconds. Auto-detects template file layout instead of applying a hardcoded column map, preventing silent mis-fills.

Security & Two Deployment Options

On-premise: entire system runs within the internal network, drawings never leave the infrastructure, AI models run internally without calling external services — recommended for documents with strict security constraints.

Cloud: ready to use without server investment, continuous feature updates, data isolated per customer — recommended for initial phases.

Common mechanisms: clear permission separation (admin pages separate from user screens), edit logs, AI access keys stored in configuration not embedded in code, job-level data isolation.

Layered Architecture

5 layers: User Interface (document processing screen · HITL review screen · template admin page) → Application Services (document intake · processing orchestration · session and log storage) → Processing & Business Rules (document-type-specific extraction · unit normalization · steel grade mapping · quantity verification) → Data & File Export (unified data model · material catalog · template library · Excel file generation) → Technology Platform.

Adding new document types doesn't require modifying calculation or quotation export logic.