Moderation queue.
Two-step pipeline: every upload runs through Nano Banana 2 as an NSFW probe (refusal = flag). Flagged users are bumped to KYC; Claude vision then verifies passport matches the account profile.
01
AI probe
Image sent to Nano Banana 2 with prompt “use this as inspiration.” Refusal = NSFW signal. Cost: $0 per check.
02
User KYC
Flagged user is asked to upload a government ID. Image moves todraft.
03
Claude review
Claude vision compares the passport against the user's account name, DOB, bio. Approve / disapprove + reason. Reason forwarded to user.
🖼️
🤖
Probe· nano-banana-2
API refused: 'cannot generate images depicting suggestive content'
🪪
🔍
Claude· claude-opus-4-7
Reviewing passport vs account...
🪪
🤖
Probe· nano-banana-2
Refused — bumped to KYC
✅
Claude· claude-opus-4-7 · conf 0.94
Passport name 'Jane Suriyathep' matches account display name. DOB consistent with stated age 25. Document features and security elements appear authentic.
🪪
🤖
Probe· nano-banana-2
Refused — bumped to KYC
✗
Claude· claude-opus-4-7 · conf 0.81
Passport surname differs from account name (passport reads 'Cruz', account reads 'Reyes'). Could be a maiden-name change — user should confirm with marriage cert or alternative ID.
🖼️
🤖
Probe· nano-banana-2
Allowed — auto-published
🖼️
🤖
Probe· nano-banana-2
Probe returned ambiguous: image describes 'editorial portrait in low light' — not refused but flagged 'low confidence' by safety classifier