Inspection reports, drafted forthe inspector to approve
The Property Inspection Pros is a home inspection firm. I built it a platform that sorts each inspection’s photos, suggests the finding and comment for each from the firm’s own library, and, once the inspector has approved everything in one app, enters the finished report into HomeGauge. Nobody copies text back and forth to ChatGPT.
56%
less time preparing each inspection report
Plates and diagrams on this page are illustrations. No client photos or data are shown.
Writing inspection reports used to take up a big part of my week. Rasheed built us an app that organizes our inspection photos and suggests comments, saving us a ton of manual work. We can review everything, make changes, and use AI to help write new comments without having to go back and forth to ChatGPT. Once we're happy with everything, the app enters it all into HomeGauge for us.
What used to take hours now takes about a third of the time, and we've used it on hundreds of inspections.
Rasheed is always finding ways to make it better, too!
90 minutes of preparation behind every report.
Each inspection ends in a report in HomeGauge, the firm’s reporting software: its sections, findings, comments and photos. Writing new comments meant copying text back and forth to ChatGPT.
HomeGauge stays
The finished report still goes into HomeGauge, the firm’s reporting software, on the inspector’s Windows computer.
The firm’s own library
Every suggested comment comes from the firm’s own comment library.
Approval first
Nothing goes into a report that the inspector has not approved.
One photo, six stages.
Follow a single inspection photo from the camera to the finished report. The system sorts, matches and suggests; the inspector approves; only then is anything entered.
- CategoryElectricalHVACPlumbingIMG 0214Electrical01
Sorted
A PyTorch classifier assigns each photo its category.
- In
- Each inspection photo
- Out
- A category
- Runs on
- A custom, fine-tuned PyTorch model
- Who decides
- System suggests; the inspector reviews
Past inspectionsIMG 0214Electrical02Matched
DINOv2 and Qdrant retrieval find visually similar photos from past inspections.
- In
- The same photo
- Out
- Visually similar photos from past inspections
- Runs on
- DINOv2 embeddings, Qdrant vector search
- Who decides
- System suggests; the inspector reviews
- FindingComment · libraryAlternativesIMG 0214Electrical03
Suggested
The two are weighted to suggest the finding and comment from the firm’s own library, with alternatives one click away; confidence scores stay visible.
- In
- The category and the matches
- Out
- A finding and comment from the firm’s library, with alternatives
- Runs on
- Classifier confidence and retrieved matches, weighted
- Who decides
- System suggests; the inspector reviews
- Two predictionsElectricalExteriorWithin a 20% marginOpenAI modelFlagged for the inspectorIMG 0214Electrical04
Close calls
An OpenAI model weighs any two predictions within a 20% confidence margin; anything still uncertain is flagged for the inspector instead of silently accepted.
- In
- Two predictions within a 20% confidence margin
- Out
- A resolved category, or a flag
- Runs on
- An OpenAI model
- Who decides
- Uncertain items go to the inspector
- ApprovedInspectorEditedNew commentIMG 0214Electrical05
Approved
Inspectors review, edit and draft new comments (with an OpenAI model) in one app. Nothing is entered until the inspector approves.
- In
- The organized photos and suggestions
- Out
- An approved report
- Runs on
- One review app; an OpenAI model for new comments
- Who decides
- The inspector
- HomeGauge
- Section
- Finding
- Comment
- Photo
VerificationIMG 0214Electrical06Entered
A FastAPI backend prepares the report and Power Automate Desktop enters it into HomeGauge on the inspector’s Windows computer; a verification pass reports any failed or mismatched entry.
- In
- The approved report
- Out
- Sections, findings, comments and photos in HomeGauge; a report of any failed or mismatched entry
- Runs on
- FastAPI; Power Automate Desktop on the inspector’s Windows computer
- Who decides
- System enters and verifies
The system suggests. The inspector decides.
The system follows a human-in-the-loop design: AI accelerates the repetitive work, while inspectors remain accountable for professional judgments and the final report.
From about 40% to 85%, on the inspectors’ own corrections.
85%
Classification and retrieval accuracy, from about 40% in December 2025
Every correction an inspector makes is captured and reviewed weekly, and feeds the model’s refinement and retraining.
Accuracy excludes new findings with no matching comment in the library yet.
Active, and reviewed every week.
The inspectors’ corrections keep driving the model’s refinement and retraining.
PyTorch/ DINOv2/ Qdrant/ OpenAI/ Power Automate Desktop/ HomeGauge/ Next.js/ TypeScript/ FastAPI/ Postgres/ MLflow/ LangSmith