Intelligence

Grounded extraction, not a guess with good handwriting.

Every figure points back at the line it came from, the loop that improves it is measured rather than assumed, spend is capped per organisation, and your drawings are processed in Australia.

Read · drawing chat

Ask the drawing anything.

A real agent with real tools. Ask it to find what you are quoting, then ask it for the lengths — and click through to the line it measured.

52740 / A-201 / CHATRunning search

Locate the balustrades and handrails on this drawing.

searchvision_locate

Three runs on A-201. Two balustrades and one stair handrail — pick one to see it on the sheet.

What's the length of the balcony balustrade?

details

BAL-1 measures 7.20 m, taken off line B–E at the sheet's own scale. Nothing was retyped — that figure still points at the line it came from.

And the internal stairs — balustrade and handrail lengths?

compare

BAL-2 is 4.80 m. HR-1 is 7.52 m — that is the RAKE length, not the horizontal projection. Quote a stair handrail on its projection and you quote it short.

A-201 · sheetBAL-1 · 7.20 m
BAL-17.20 m

Grounding

Four things that keep it anchored to the page.

The difference between a model that reads a drawing and a model that writes plausible numbers about one.

  • Deterministic first

    Where a value can be measured or read directly off the page, it is. The model is asked to extend that grounding, not to originate every figure from scratch.

  • Item identity across pages

    A run named on a schedule and shown again on a detail sheet is tracked as the one item it is, so nothing is silently doubled.

  • Document hierarchy

    Floor plans, elevations, sections, schedules and addenda are read in a declared order of authority, not as one undifferentiated pile of pages.

  • Every figure cites its source

    Each extracted value traces back to the page and region it came from, so a reviewer can go to the source instead of re-reading the whole set.

The flywheel

A loop that closes, not a pipeline that runs forward.

Every confirmation and every correction becomes a training label, and each stage below names a subsystem that ships — the list is asserted against the repo, so it cannot outlive the machinery it describes.

  1. 01

    You mark up

    An estimator marks the runs on a drawing, the way they always have — because the marks are the job, not a chore added for the model.

  2. 02

    Marks become labels

    Every confirmation and every correction is captured as a training label, tied to the exact geometry it came from.

  3. 03

    The studio catches the hard ones

    Where the model was unsure, a human confirms or corrects the element type — nothing is written without an explicit confirm.

  4. 04

    Models retrain

    The ranker learns which candidates estimators actually keep, and the detector retrains on the located boxes.

  5. 05

    It proposes more, and better

    Auto-marking puts up more candidates, and the ones it puts up are ones you are more likely to accept.

  6. 06

    You correct less

    Fewer edits per drawing, and the ones left are the genuinely ambiguous ones — and each of those corrections is the next training label, which is where this started.

…and that correction is the next label. The loop closes.

Cost & routing

Governed spend, routed to the right model for the job.

A per-tenant cap enforced before the call, and the strongest model only where the work needs it.

  • Per-org daily spend caps

    Each organisation has its own daily AI budget, enforced before a call is made. One tenant running hot cannot spend another tenant's budget.

  • The hardest step gets the strongest model

    Reading dimensions off a technical drawing is the hardest thing this platform does, and it runs on the best spatial-reasoning model available to us — reviewed as better ones ship, rather than fixed at whatever was current on the day we built it.

  • The routine step runs on a cheaper one

    Turning confirmed dimensions into quote lines is structured work, so it runs on a fast, inexpensive model — and escalates to the strong one automatically when confidence is low. You are not paying frontier rates to format a line item.

Data residency

Your drawings are processed in Australia.

Architectural drawings are commercially sensitive, and for a builder working on government or hospital projects, where they get processed is a procurement condition rather than a detail.

Extraction runs in-region

Drawing analysis runs on models hosted in Australian regions. That is the shipped default, not a configuration you have to ask for.

Drawings and job data stay here

Documents, extractions, quotes and job records live in Australian infrastructure for the life of the account.

Your drawings are not someone else's training set

What the platform learns from your corrections improves your organisation's own results. Nothing is pooled into a shared model across customers.

Point it at a set you already know the answer to.

Grounding is only a claim until you check it against a job you priced yourself.