AI document search for a domain no model has ever seen.
Arcmaze
Query time reduction
Query routing precision
Saved per architect per week
Pilot users onboarded
The Situation
Construction drawings encode information in symbols, line weights, spatial relationships, and tiny annotations. No public dataset of annotated construction drawings exists. No AI model has been trained on them. A misread dimension isn’t a search error. It’s a structural risk.
The Problem
The daily reality: finding a specific detail buried in 50–75 drawings means opening sheets one by one and scanning. 4 to 6 hours per query. Standard AI approaches failed in testing. Full-sheet LLM processing caused hallucination. Plain OCR missed most of the useful information. The problem wasn’t model quality. It was architecture.
What We Built
Multimodal RAG platform with a query routing layer at its core
Auto-categorises documents by discipline at ingestion (mechanical, electrical, plumbing, structural)
Natural language query routing. Narrows retrieval context before the model processes anything
Built through direct collaboration with a practising architect as domain teacher
Zero prior domain knowledge. Every design decision earned from first principles
The Outcome
Currently piloting with 10–15 users across multiple large-scale projects. Query time down from 4–6 hours to under 2 minutes. Routing precision at 70–85%. Estimated 3–5 hours saved per architect per week. These are pilot-stage results. The technical proof of concept holds.
“We started with zero domain knowledge. Every design decision traces back to learning how architects actually think.”
- Pixeldust Team
Have a project in mind?
Pixeldust