Triple-AI Verified Takeoffs — three models, one number you can sign
Why running every drawing through Claude, Gemini, and a self-hosted YOLO catches what single-model AI takeoffs miss — and how ORKSTRA routes by criticality.
A single AI takeoff is a confident guess. A Triple-AI takeoff is a checked answer. The difference matters at tender pricing, where a 3% optimistic count eats a job's profit; and it matters at IPC time, where a missed line delays payment by weeks.
ORKSTRA runs every drawing through three independent vision models — and that's not a marketing line, it's a routing architecture.
The three models, and what they each do
Claude is the reasoning layer. It reads the specifications, the BOQ context, RFI threads, and contract clauses. When a recessed light callout says "Type B with dimmer per Architectural Spec 16-2.3," Claude is the model that goes and reads section 16-2.3.
Gemini is the drawing-vision layer. It reads PDF and DWG sheets natively. It extracts dimensions, callouts and counts even from scanned construction drawings where OCR alone falls over.
YOLO is the symbol layer — self-hosted YOLOv8 with a 52-class construction catalog. It finds the diffusers, the fire heads, the junction boxes, the doors. Then it verifies the counts the other two models claim to see.
Why three is better than one
Pick any one of those models in isolation and it will fail confidently. It will hallucinate a count, or miss a recessed light, or read a 1:50 sheet as 1:100. You won't know — until the QS catches it in week 6, or the consultant catches it in week 8.
Three models in parallel turn silent failure into visible disagreement. ORKSTRA's Triple-AI router compares outputs from the three; anywhere two of them disagree by more than a tolerance, the line is flagged for human review. The estimator reviews the disagreements (typically 5-15% of items) instead of every line.
Cost-aware routing
You don't need triple-AI on every line. Routine material classifications go through Gemini alone for fractions of a cent. Drawing-to-BOQ mapping goes through Claude + Gemini for a few cents. Critical contractual numbers — tender bids, IPCs, regulator submissions — go through the full ensemble plus a senior QS sign-off.
The router decides per item, based on criticality tags ORKSTRA learned from your tenant's history.
What it means at scale
A typical UAE BOQ runs 800–2,000 items. Pre-Triple-AI, a senior QS spent ~28 hours per BOQ on takeoffs. With Triple-AI handling the routine 85–90% and the QS focused on the 10–15% disagreements, the same BOQ closes in ~5 hours. Same quality. Higher confidence. The output is a number you can put your name on.
Try it on one of your own drawings — book a 20-minute demo and we'll run Triple-AI live on whatever you bring.