Floor Plan Detection/FastAPI/Python/Computer Vision

I’m building an end-to-end service that takes architectural floor-plan photographs/PDFs and returns a structured JSON map of rooms, including room polygons, labels and confidence scores.

I’m looking for a hands-on Python/FastAPI + Computer Vision engineer who can understand and debug an existing codebase, trace the complete image-processing flow, and significantly improve room detection accuracy.

Key Requirements

Strong Python + FastAPI

-Hands-on production experience with Python and FastAPI
-Comfortable debugging and improving existing Python code
-Experience with APIs, file uploads, Pydantic, JSON and async processing

Computer Vision — very important

-Strong practical OpenCV experience
-Image preprocessing, thresholding, contours, connected components, morphology, line/edge detection, perspective correction and segmentation
-Experience extracting regions/polygons from images
-OCR integration is a plus
-Ability to diagnose why a room was missed and improve the underlying CV pipeline

Difficult / unaligned floor plans
A major challenge is detecting rooms that are:

-Rotated or skewed
-Non-rectangular / L-shaped / irregular
-Not aligned with the image axes
-Partially obscured
-Surrounded by broken/faint wall lines
-Affected by perspective distortion, shadows or noise

Experience with floor plans, document analysis, segmentation, OCR, or similar geometry-heavy CV problems is a strong plus.

Other skills

-Able to read and understand TypeScript
-Comfortable with Bash/Linux/Git
-Experience with Anthropic Claude API / LLMs is a plus

Claude will be used mainly for higher-level reasoning such as room-label normalisation, resolving OCR ambiguity and understanding room relationships. The core room detection should come from a robust CV pipeline.

Deliverables
-Clean, commented Python codebase in Git
-FastAPI /detect endpoint supporting PDF/JPEG/PNG
-JSON output with room polygons, labels and confidence scores
-README with setup and debugging/training instructions
-Small demo/test dataset
-Accuracy report
-Improvements to the CV detection pipeline

Target
-≥90% room detection/identification accuracy on supplied test photographs
-Approximately