The change was on the drawing. The cloud around it was not.
Design and construction run on documents that change faster than anyone can check them. ZoraAGI puts agents on exactly that: Zora BluePrint AI™ reads two revisions and computes what actually moved — then flags every change nobody clouded. Zora BluePrint IQ™ takes an estate of project issue logs, written once and never read again, and turns it into memory the next project can use. Both run as agents on the Zora AGI™ OS, inside your network.
A revision cloud is a promise. It is not a guarantee.
When a sheet is reissued, the clouds are meant to mark what moved. In practice a dimension chain is edited and never clouded, the sheet is released, and the change reaches fabrication unannounced. The cost of that surfaces weeks later, in steel that does not fit. Meanwhile every issue log that records these failures is closed out and never queried again — hundreds of projects solving the same detail from scratch, each one alone.
How drawings are checked today
- Manual overlayTwo prints on a light table, or two PDFs toggled by eye. Accurate right up until attention runs out.
- Pixel-diff toolsFlag every anti-aliased edge and every re-scaled title block, so the one change that matters drowns in noise.
- Trusting the cloudsWhich finds precisely the changes somebody remembered to cloud, and none of the rest.
- Issue logsWritten for the project that raised them, closed at handover, and never read across projects again.
How agents check them
- Vector set-differenceText and geometry compared as data rather than pixels. The change count is computed, not estimated.
- Cloud reconciliationEvery computed change is tested against the revision clouds; anything falling outside one is raised as documentation risk.
- Visual agents on the cropA spatial auditor and a code-compliance checker read the changed region at full resolution and say what it means.
- Cross-project recallEvery issue ever logged becomes retrievable evidence for the sheet somebody is reviewing right now.
The numbers come from mathematics. The models are only allowed to explain them.
Upload two revisions. Get an audit-grade change register.
An engineer signs in on the LAN, submits a sheet pair, watches the pipeline run, and reviews a report where every difference is listed, marked up on an overlay, explained in plain language, and — where no cloud encloses it — flagged in red. Approve it, or send it back with comments.
Completeness & quality audit
Type-specific rule packs check a single sheet before it is issued — missing dimensions, absent callouts, unresolved references, deviations from the standard you seeded the knowledge base with.
Revision delta
Exact set-difference over extracted vector text and geometry, reconciled against the revision clouds. Undocumented changes are flagged as risk rather than buried in a list.
Registers both sheets into one coordinate space, then computes the exact difference in text and geometry. No model involved, and none needed.
Detects revision clouds on the sheet and tests every computed change against them. What sits outside becomes the headline finding.
Reads the cropped change region at 300 DPI and describes what actually changed, in the language an engineer would use in an RFI.
Cross-checks the change against standards and golden drawings retrieved from the knowledge base, and cites what it checked against.
Three artefacts come out of every run: an overlay-marked PDF, an Excel change register, and structured RFIs ready to push into Autodesk Construction Cloud or Procore. A supervisor sequences the agents; a critic and a confidence gate decide what a person needs to look at.
Runs on one workstation. The whole pipeline deploys on a single GPU box inside the client network, with automatic CPU fallback if the GPU is busy or absent. Local models only, LAN-only access, and no per-seat AI subscription anywhere in the stack.
One project's issue log is a to-do list. Three hundred are a dataset.
Design comments, clashes, quality defects and commissioning items get logged diligently on every project, and then never looked at again. The same waterproofing detail fails on project 41 the way it failed on project 7. The duct-and-beam clash is investigated from scratch each time. How it was fixed retires with the closeout report. BluePrint IQ makes that estate queryable — and puts the answer in front of the reviewer while the issue is still open.
Clusters issues across projects on four signals at once — how they look, how they read, where they sit, and what element they touch — into a curated library of known patterns.
The moment an issue is raised, agents retrieve similar past issues together with their resolutions and outcomes, and post an evidence-linked suggestion card back into the tool within seconds.
Pins, sheets, attachments and comment threads are normalised into one issue schema and a knowledge graph, with provenance recorded on every extracted field.
Visual fingerprinting matches the sheet under review against historical sheets across the estate, then projects past issue pins onto it as a heatmap.
Where it plugs in. BluePrint IQ reads an Autodesk Construction Cloud estate through the Issues, Data Management and Model Derivative APIs, with webhooks driving real-time suggestions. The data fabric underneath is backend-agnostic, so the same system runs on-premises or air-gapped where the estate cannot move.
Send us a revision pair. We will send back the register.
The fastest way to judge this is on your own sheets. Give us a Rev A and a Rev B, and we will return the computed change register, the overlay markup, and the count of changes that were issued without a cloud — then walk you through how each one was found.