RevRadar uses AI vision and language models to assist with reviewing revisions between construction drawings. This document explains, in plain English, what the tool can and cannot do, and what is expected of you as a user.
You acknowledge this policy when you create an account.
1. What the Service is
RevRadar performs AI-assisted change detection between two revisions of the same drawing set. The Service:
- Renders pages of each PDF as images.
- Compares the rendered pages and identifies regions that appear to have changed.
- Sends crops of those regions, and rendered title-block sections, to AI models for triage and summarisation.
- Produces a report of changes identified by the AI scan, grouped by severity, alongside source-of-truth crops you can verify by eye.
2. What the Service is NOT
The Service is not a substitute for a qualified human review. Specifically:
- It does not guarantee that every change between revisions has been identified. AI models can miss changes, mis-classify them, or flag changes that are not actually present (false positives).
- It is not a certified or accredited review. No professional sign- off or stamp is implied by a RevRadar report.
- It does not absolve any architect, engineer, project manager, superintendent, or other qualified person of their duty to independently verify drawings and make their own judgement before acting on findings.
- It is not a structural, code-compliance, or constructability check.
3. Your responsibilities
By using the Service, you agree that you will:
- Treat every RevRadar report as a review aid, not a final determination of what has changed.
- Independently verify every flagged change before acting on it, using the original source drawings.
- Independently look for changes the Service may have missed, particularly in areas the Service was unable to scan, in regions flagged as low-confidence, or on pages that failed to process.
- Not rely on the Service alone for any decision that has safety, contractual, or regulatory consequences.
- Ensure that any human reviewer you delegate this work to is qualified to perform construction drawing review.
4. How AI is used
- Vision: rendered images of crop regions and title blocks are sent to large language models for classification and summarisation.
- Models we use: see the Privacy Policy for the current list of sub-processors. The model name, model version, and prompt version used to produce each report are stamped on the report and on the job detail page so you can audit which version produced which output.
- No training: we do not contribute your drawings to model training. Sub-processor settings are configured to disable training on customer data wherever the provider supports it.
- Hallucination risk: AI models can fabricate plausible-looking output. The Service tries to mitigate this by anchoring summaries to specific image crops and surfacing the underlying crops in the report, but it cannot eliminate the risk. Always check the source imagery alongside the AI-generated summary text.
5. Limitations of the Service
The Service may, without notice, fail to identify or correctly classify a change because of:
- Drawings that are scanned, rotated, of low resolution, or contain raster-only content.
- Differences that are present but visually small relative to the page (e.g. a single dimension change on a busy plan).
- Differences that require domain context the AI does not have (e.g. specification or specification-cross-reference changes).
- PDFs that are corrupt, password-protected, or otherwise unable to be processed.
- Service outages, model provider outages, or rate-limit conditions that cause processing to be partially completed.
When a page fails to process or the Service is uncertain about a region, this will be surfaced on the report. The absence of a flagged change in any given region is not a guarantee that no change exists in that region.
6. Versioning and re-acceptance
We may materially update the way the Service detects changes (new models, new prompts, changed pipelines). When we do, we will either:
- Stamp the new model / prompt / lib version on every subsequent report, so you can identify which version produced which output; or
- If the change is material enough to affect what users should rely on the Service for, ask you to re-acknowledge an updated version of this policy before continuing to submit jobs.
7. Audit history
We retain an audit record of:
- Each detection run (input file hashes, model version, prompt version, output hashes, timestamps).
- Each acknowledgment of this policy and our Terms of Service (timestamp, IP, user-agent, version).
- Each time a report is viewed or downloaded.
These records are retained beyond the retention windows that apply to the underlying drawings and reports — see the Privacy Policy for details.
8. Contact
Questions about this policy or the Service: lachie@asmblr.com.au.