Cortex AI
Model layer automating quantity takeoff
2024 to 2026 · Technical Product Manager
- Defined the AI layer inside Geometra before engineering picked it up: the model approach, the fine-tuning strategy, and the dataset creation plan the training rested on.
- The dataset was where the product decisions lived. What goes into it defines what the system treats as correct, and in takeoff there is no neutral answer: whether a shared wall counts once or twice, how an ambiguous hatch pattern resolves, what a surveyor would have done. Those are pricing decisions disguised as labelling decisions, which is why they did not belong to engineering.
- A wrong output becomes a wrong number on a customer quote, so we gated release on per-class precision against surveyor-verified drawings, not on aggregate accuracy. Averages hide exactly the failures that cost money.
- Fine-tuned SAM 2 to auto-annotate architectural drawings and emit XFDF straight into the quotation system. In production with customers.
Built with
- AI Product
- Fine-tuning Strategy
- Dataset Design
- SAM 2
- Computer Vision
- Model Evaluation