Construction · AI Consulting & Automation · Software Consulting & Engineering
AI Revolution in Construction: Weeks Cut to Minutes
RITE NRG partnered with Swedish founders to develop an AI-powered SaaS that reduces construction installation estimates from weeks to minutes. Through a three-month discovery phase, we trained a custom AI model reaching 80% accuracy in recognizing water pipes from 2D drawings, laying the foundation for a market-ready product.
80%
Accuracy recognizing water installations from 2D drawings
3 months
Discovery phase from feasibility study to trained model
Full clarity
MVP scope, timeline and budget defined for the next phase
Starting situation
Two investors with deep construction experience identified a critical market gap: preparing installation offers and take-offs from 2D drawings took weeks of manual, repetitive work. The primary challenge was an AI solution able to analyze 2D architectural and technical drawings accurately enough to generate cost and material calculations, starting with water installations. Standard models such as ChatGPT and Gemini proved insufficient for the specialized task. The founding team also lacked a technical leader, product development know-how and any in-house development capacity.
What we advised
We proposed a phased approach beginning with a focused discovery phase to validate whether AI could support construction offer preparation before committing to full development. A feasibility study assessed whether existing models such as YOLOv8 and Detectron2 could identify and segment technical installations on PDF drawings, giving the founders a cost-effective read on the technology's potential. We also acted as a fractional CTO, advising on product direction and market strategy rather than only writing code.
What we built or changed
Over three months we prepared and processed technical drawings in PDF form, annotated images and trained a UNet with MobileNetV2 encoder to identify water installations. Weekly demos and status reports kept the founders in control of direction. As part of the #riteway, we ran a hands-on product design workshop in Wroclaw to define user needs, product features and the MVP scope alongside the founders.
Architecture, data and integration
The stack was built on PyTorch 1.13.1 with torchvision 0.14.1 and CUDA 11.6. Image and mask manipulation used PIL and NumPy with a custom tiling system for large drawings (1024px tiles with 256px overlap). Training used the Adam optimizer with a ReduceLROnPlateau scheduler, CrossEntropyLoss and early stopping. After evaluating Detectron2 (Mask R-CNN) and SegFormer, UNet with an ImageNet pre-trained MobileNetV2 encoder was selected, with skip connections and six segmentation classes at 1024x1024 px.
Engagement and commercial model
A bounded, fixed-scope discovery phase with a fractional CTO and a dedicated model training team, followed by a costed MVP proposal the founders could approve independently.
The #riteway in action
Weekly demos, transparent status reporting and a joint product design workshop kept business goals and technical work aligned. Our overdelivery mindset meant advising on commercial viability and go-to-market as well as delivering the model.
Delivery timeline
Three-month discovery phase, followed by a defined MVP roadmap, timeline and budget.
Handover and ongoing operation
The founders received a trained model, a documented MVP scope with effort estimation, timeline and budget breakdown, and a strategic partner able to build the product with them.
Discuss a Similar Challenge
Bring us the system, the constraint or the outcome you need. We will explain what a responsible path forward looks like.