AI-first measure + verify; interactive/dynamic takeoff¶
| Stage | Building |
|---|---|
| Status | Partial |
| Design status | In progress |
| Estimate | XL |
| Confidence | Low |
| Linear | PIN-412 ↗ PIN-295 ↗ PIN-439 ↗ |
| Linear status | Done · +1 backlog |
| Cycle | C17 |
| Design | Design ↗ |
| Linear epic | Drawings, Documents & Takeoff |
| Module | Takeoff Engine ↗ |
Priority: High · Route: /bids/:bidId/pdf-takeoff/:bidDocumentId
Scope¶
AI-driven takeoff: the estimator runs a per-system AI detection pass on the PDF drawing (auto-scale read validated against the 2×2 ceiling grid, vectorized-PDF input), reviews shaded matches through a per-system yes/no gate, and confirms to auto-fill the takeoff sheet (CSI code, system, size, qty, unit) and the estimate. A color-coded system overlay (by pressure class for ductwork, by system type for piping) lets the estimator visually verify detections instead of reading a raw list; unidentified items are flagged in an alert color. (Source: PIN-412, PIN-295, PIN-439.)
Acceptance criteria¶
- "Run AI Pass" per system, with a 4-state flow: Idle → Running (live progress) → Reviewing (detected symbols, yes/no per match) → Confirmed
- Confirming a pass auto-fills the takeoff sheet row-by-row (CSI code, system, size, quantity, unit) and triggers estimate line-item auto-fill
- AI reads the drawing's scale and validates it against the 2×2 ceiling grid before detecting
- Detected matches shade the drawing; estimator corrections (shading a missed symbol) train the model forward
- Color-coded overlay: ductwork shaded by pressure class (shade variants of one color), piping shaded by system type (CHW/CW/HW/etc.), each toggleable on/off
- Items the AI can't identify render in a distinct alert color and route to the Concordance review tab
- Input is vectorized PDFs; accuracy tuning depends on a clean training drawing corpus and the self-hosted model track (tracked as a separate open question)
Conor's comments¶
This requires a lot of design discussion. This is a tough build and has to be PDF focused. We need to build training fiedls and train ai on detection and estimation.
Notes¶
Your #1. Base pipeline shipped; accuracy is unstarted research
Open question for Conor¶
Clean drawing corpus (PIN-591) + self-hosted model sign-off (PIN-590)