Features

From tender PDF to defensible schedule

StructCount is a pipeline, not a magic button. Each stage does one job well — and you can see, and override, every decision it makes.

Stage 1 · Triage

Every sheet is understood before it's counted

The moment a tender PDF lands, each page is classified: true vector sheets (CAD-exported geometry), scanned raster sheets, and sparse pages like cover sheets or schedules. Counting strategy follows the sheet — vector sheets get exact geometry matching, scans get image template matching, and schedule pages don't waste anyone's time.

A 700-page tender document doesn't grind the system: sheets are analyzed when you first open them, so upload stays fast and you start reviewing the sheets that matter.

Stage 2 · Legend

The legend is the contract — so we read it properly

Symbols on an M&E drawing mean whatever the legend says they mean. StructCount finds the legend or key on the sheet, extracts each row's symbol cell and label, and treats every row as a symbol class to count. Text drawn as outlines (a common CAD export quirk) is recovered with OCR; enclosed tags like the letter inside a detector circle become part of the symbol's identity.

When the legend is a raster image, lives on a separate cover sheet, or simply isn't found — drag a box around it anywhere and a vision model reads the rows. The AI only reads; every count still comes from the matching engine.

Stage 3 · Matching

Exact geometry, not pixel guessing

Each symbol is matched as a constellation of vector paths: the engine anchors on the symbol's most distinctive stroke, then verifies the full arrangement — rotation-invariant, scale-aware, any angle. A smoke detector and a sounder are both circles; what separates them is the exact geometry inside, and that's precisely what the matcher checks.

When several legend symbols share structure, every candidate location is assigned to exactly one owner — the class whose template explains the most of the drawing's actual ink at that spot. No double counting, no orphan look-alikes.

Stage 4 · Review

A queue built to be emptied

Precision comes first: only distinctive, high-confidence matches auto-confirm. Everything else — partial matches, dense overlaps, drawing-edge oddities — waits in a review queue with a zoomed thumbnail per item. Work it entirely from the keyboard: j/k to move, y/n to decide. Accept or reject a whole class in one action when the engine's calls are obviously right.

An optional AI adjudicator pre-sorts the queue, marking items it would confirm or reject with its reasoning — you stay the judge. And when the legend missed a symbol family, place a few by hand and Find Similar recovers the siblings from your placements.

Stage 5 · Export

Numbers that survive scrutiny

Finalize a sheet and export the schedule as XLSX or CSV — class, count, sheet reference. Until then, every number in the panel is live: click it and the viewer highlights each counted location on the drawing. If a count is ever challenged, the evidence is one click away.

Re-running a sheet produces identical results. Deterministic counting means your audit trail is the drawing itself.

See it on your own drawings

The demo that matters is your tender package. Free while in early access.

Get started free