Text across an image or a cropped person can make an outfit hard to judge. We use Claude to help classify photos and find images that need a closer look before they enter Dait’s outfit library.
This is an internal photo review workflow. Claude reads the images and suggests classifications or quality flags. Our review and publishing tools handle what reaches the library.
Classify the incoming photos
Claude reviews a photo and returns a classification. The saved result helps us organise incoming images and identify those that need review.
Review the uncertain ones
We hold photos that need a decision. The reviewer can keep an image, correct its description or leave it on hold. A model flag alone does not decide that a photo is unusable.
Update the library
Our publishing tools apply the accepted results to the photo library used by Dait. We also review existing photos rather than only checking new arrivals.
Claude’s role here is classification and quality review. Virtual try-on uses separate image-generation tools. We have not verified Claude-powered outfit recommendations inside the consumer app.
What we’ve used it for.
These are two recorded development and photo-operations batches. They describe the work behind the library, not how many people use Dait.
Classification responses, review decisions and publishing reports are kept in our internal records.
Internal work records from October 10, 2026. Counts refer to photos, not users or app installs. This is a written account of past runs, not a live demonstration.
Batch
Recorded result
80 incoming photos
After classification and intake checks, 67 photos were published and 13 remained on hold. The published group included seven photos whose copy was corrected.
14,323 existing photos
Claude flagged 128 photos for a closer look, including text overlays, lettering issues and faces that were not visible. A flag was a review suggestion, not a final rejection.
In a human review of 33 photos, 29 were kept and four were marked to hide. Two of those four were newly excluded; the other two were already excluded. This is why we keep human review in the process.
One classification, followed through.
Input record
c00004 · October 10 intake
Saved Claude classification
Black halter-neck long dress, beige sandals. Summer, clear weather, 24–31°C. Marked usable.
Intake check
Ready, with no hold reason. Published in the intake batch.
Library record
7a0b139c-5429-41ed-8699-c62b59632aae The original file hash matches this record, and its display image was reachable.
This is a summary of the saved, parsed classification and publication records, not a replay of the API call. The intake check was automatic; there is no individual human approval record for this photo. Its appearance on an actual app screen still needs to be captured.
What these records show
They show that Claude has been used in Dait’s internal photo classification and review work. They do not establish review accuracy, customer traction or a public app launch. A walkthrough linking one photo’s saved response, review decision and actual app screen is still being prepared.
The next review pass.
We want to check which visual issues matter when choosing an outfit, and where Claude’s flags agree or disagree with a person’s review.
We’ll compare missed problems, unnecessary flags, response time and cost on photos we can use. The aim is to improve the review process before expanding it.
Made by BeeK.
BeeK develops apps in South Korea. We handle product design, development, testing and operation ourselves.
Our business opened in October 2024. We fund development without outside investment.
Registered business
비크 (BeeK) · Sole proprietorship
Business opening date
Representative
손태섭
Business registration
386-71-00582
Location
South Korea
Funding
Bootstrapped · No external investment
We also use Claude Code to review code and find errors. We run and test proposed changes before putting them into the app.