FOSVED
← All projects

Alpha

CAD Reconstructor AI

Turns a photo or scan of a technical drawing into a ready-to-edit DXF file with a real scale

3 of 5 done · In progress: Green build and the 0.8.0 release

Screenshots and mockups

  • Drawing upload: file and scale selection
    Drawing upload: file and scale selection
  • Processing: the nine steps of the reconstruction pipeline
    Processing: the nine steps of the reconstruction pipeline
  • Manual editor: detected lines over the source plan
    Manual editor: detected lines over the source plan
  • Scale calibration from a reference dimension
    Scale calibration from a reference dimension
  • Settings: connecting your own AI provider key
    Settings: connecting your own AI provider key

A web service that takes a photo or scan of a 2D technical drawing (for example a floor plan) and returns a clean DXF file that opens in AutoCAD and similar tools. The value is not in tracing pixels but in recovering meaning: it reads dimension numbers, derives the real scale from them and redrafts the plan from scratch — walls with thickness, doors, windows, room labels, dimension chains. Two engines are available: classic computer vision for clean scans and a vision-language model for photos of paper or hand-drawn plans. The output follows the Russian and Belarusian drafting standards: frame, title block, coordinate axes, room schedule.

What exists

  • Two reconstruction engines with automatic fallback to the classic one if the model fails
  • The scale is derived from recognised dimensions, and the source of the scale is recorded in the report
  • Drafting per ESKD, GOST and SPDS — frame, title block, axes, hatching, room schedule
  • Stitching of 2 to 12 overlapping photos of a large drawing into one image
  • Phone-camera capture directly from the browser
  • An online editor with a tracing underlay, layers and scale calibration against a known segment
  • A quality report with an honest verdict — if the scale is unreliable, delivery is blocked until manual calibration
  • Pay per file, the first one free

Highlights

  • The scale is not guessed but derived with proof — dimensions are grouped into chains and accepted only if the segments add up to the overall size
  • The output follows national drafting standards rather than being merely geometrically correct
  • A ban on "successful garbage" — if the result is unconvincing the system honestly flags it as needing review

Not there yet

  • There is no public demo or deployment
  • No tagged releases yet (0.8.0 appears only in project files)
  • Only flat plans are supported — no 3D, assemblies, sections or mechanical drawings

Progress

  1. End-to-end path from upload to DXF downloadDone
  2. Drafting per ESKD, GOST and SPDSDone
  3. Large-drawing stitching and phone captureDone
  4. Green build and the 0.8.0 releaseIn progress
  5. Public deployment and first usersPlanned

Next step: The build is green again. What remains is tagging release 0.8.0.

Stack

  • Python
  • TypeScript
  • FastAPI
  • SQLAlchemy
  • Celery
  • OpenCV
  • Shapely
  • ezdxf
  • PaddleOCR
  • React
  • Vite
  • Fabric.js
  • PostgreSQL
  • Redis
  • Docker Compose
  • nginx
  • GitHub Actions
  • Vision-language model through an OpenAI-compatible interface
  • PaddleOCR and EasyOCR
← All projects