Case study / Feb.2026 / Dev & AI / Client: BTP Consultants
AnalyzTech
AnalyzTech is an AI system that helps technical engineers and compliance controllers verify floorplans and large construction documents. It reads plan PDFs, extracts and classifies rooms, and presents the results in a review interface so verification takes less repetitive manual work.
01 / Brief
Solution
The pipeline starts from a plan PDF. Kreo extracts walls and doors, YOLO detects furniture and open doors, and room extraction turns those elements into room polygons. A vision-language model classifies each room, post-processing cleans the result, and an AI Agent runs verification actions on the structured output.
I built the FastAPI backend, model pipeline, polygon-review interface, AI Agent verification skills and AWS deployment end to end. The backend routes project context to the agent, while the frontend lets engineers inspect extracted polygons and trigger AI-assisted checks from the plan view.
02 / Constraints
Problem
- Building a reliable floorplan reading workflow from scratch (dataset, detection, classification, extraction)
- Improving robustness on large documents and noisy plan structures
- Reaching high accuracy for room classification and room-mask detection
- Reducing latency so users can run verifications without long waits
03 / System
Architecture / workflow
- Pipeline: PDF input -> Kreo wall/door detection -> YOLO furniture/open-door detection -> room extraction -> VLM room classification -> post-processing -> AI Agent verification
- Full FastAPI backend with integrated verification workflows and project context routing to the AI Agent
- Frontend integration to visualize polygons on plans and trigger AI-assisted controls
- Classification and detection accuracy optimization through iterative model and post-processing tuning
- Production deployment on AWS with scalable API infrastructure
AI-assisted verification supports engineers; it does not replace their professional judgment. This case describes the built plan-PDF, classification and verification pipeline, not performance on other clients' plans or workflows.
04 / Evidence
Results / impact
These figures are project-context outcomes. Detailed baselines and calculation methods are not public unless stated, and the results do not establish performance for a different workflow.
- 80% time saved on verification workflows
- +50% productivity gain
- 99% classification precision
- 96% room-mask precision
05 / Gallery