AMELA | EstiScan CASE STUDY
Open the demo →
Construction & Drawing Analysis × Hybrid AI

From a pixel-only TIFF to
mm-level takeoff you can measure.

This isn’t “hand it all to a VLM on full-auto.” Deterministic methods handle coordinates & dimensions, AI handles meaning & Japanese, and risk areas are verified by humans. We compute the takeoff from traceable structured data, not from estimates.

Try the 5-step pipeline → See how it works
Drawings & takeoff
100%
Traceable structured output
THE PROBLEM

Scanned drawings hold only “pixels”

Left to a VLM alone, dimensions become “estimates,” and miscounts and hallucinations propagate into the takeoff (= cost).

Becomes an estimate
Dimensions become “roughly”

A pixel-only TIFF leaves lines, text, and parts unreadable by machines. With a VLM alone, dimensions become estimates.

Errors propagate
Amplified by unit price × quantity

Miscounts and hallucinations grow into cost errors through unit-price × quantity takeoff.

Can’t be explained
Black box

You can’t trace or audit “why that value.” You can’t explain the basis for the cost afterward.

THE SOLUTION

A “right tool for the job” hybrid in 5 steps

▣ Deterministic= coordinates & dimensions (mm) ◆ AI= meaning & Japanese 👤 Human= review of risk areas

Measure dimensions “deterministically”

Preprocessing & vectorization convert pixels into line segments with start/end points and coordinates. Dimensions are fixed geometrically, not estimated by AI.

OSS-centric, commercially clean

Built around commercially usable OSS such as OpenCV / DeepLSD / PaddleOCR / Detectron2 / Qwen2.5-VL, avoiding vendor lock-in.

INTERACTIVE DEMO

Try the three scenarios

Click to experience the actual processing flow (mock data) right in your browser.

THE IMPACT

Not estimates, but measurable takeoff

mmunit
Geometry-based dimensions & coordinates
5steps
Hybrid multi-stage processing
100%
Traceable & auditable
OSS
Commercially clean setup
Tech stack (OSS-centric)
OpenCV DeepLSD PaddleOCR Detectron2 Qwen2.5-VL ezdxf / IfcOpenShell