Privacy Policy

Effective date: October 11, 2026

PixelProbe ("we", "us", "our") operates the PixelProbe mobile application ("the App"). This policy describes what data we collect, what we do with it, which third parties receive it, how long we keep it, and the rights you have over it. We aim to be concrete and verifiable rather than vague.

1. What We Collect

Account data. Your email address (required for registration). If you sign in with Apple or Google, we also store the provider's stable user identifier (Apple's sub claim or Google's sub claim) and the email address shared by the provider. We never see your Apple ID or Google account password.

Authentication state. Hashed password (bcrypt) for email/password accounts; short-lived 6-digit verification and password-reset codes (15-minute TTL each); a session token version used to invalidate refresh tokens on logout.

Images you submit for analysis. The App resizes each photo on your device to at most 1024 px on its longest side before it leaves the phone; the full-resolution original is never uploaded. From what arrives, the server computes a perceptual hash (pHash) and a SHA-256 hash, and writes two JPEG copies to storage: a 200 px thumbnail and a 1200 px (max side) display copy. The uploaded byte stream itself is held in memory for the duration of the analysis and then discarded.

Scan results. The verdict (authentic / uncertain / AI-edited), the numeric confidence, the identifier of the model that produced it, the written explanation, the coordinates of any region the model marked as suspect, and timing metadata, linked to your account.

Subscription state. Whether you are on the free tier or Pro Monthly, your current credit balance, your Pro 30-day scan window counters, and which pack purchases have been applied. We do not see or store payment card details — Apple processes all payments.

Server-side operational logs. Standard request logs at our hosting provider (Railway) include IP address, request path, response status, and timestamps. These are not connected to your account in our application database but may be retained by the hosting provider per its policies.

2. How We Use This Data

We do not use your data for advertising, ad measurement, behavioural profiling, or for sale to data brokers. We do not use your images to train any model of our own, and we do not grant anyone the right to train on them.

3. Third Parties That Receive Your Data

The list below is concrete: each entry states what we send, where the recipient is based, and whether the data can be linked to your identity on the recipient's side.

For image analysis

For image hosting

For billing

For authentication

For email delivery

For hosting

We do not currently use crash-reporting or analytics services. The Sentry SDK is present in the codebase but is not configured with a destination, so nothing is sent anywhere. If we enable any such service later, this section will be updated before the change takes effect.

4. How Images Are Handled End-to-End

This is the most sensitive part of the data flow. The full sequence:

  1. The App resizes the photo on your device to at most 1024 px on the longest side and uploads that copy over HTTPS to our backend. The original stays on your phone.
  2. The server computes a SHA-256 and a pHash of the uploaded bytes in memory and stores both on the scan record.
  3. If some earlier completed scan has the identical SHA-256 — the same file, byte for byte — its verdict and explanation are reused and nothing is sent to any model. This match is not limited to your own scans: if another account previously analysed exactly the same file, its result is what you get. The other account's identity, email and history are never exposed, and no image moves between accounts.
  4. Two compressed JPEG copies are written to Supabase Storage at thumbnails/<scan_id>.jpg (200 px) and images/<scan_id>.jpg (1200 px max side).
  5. If there was no exact match, the uploaded bytes are sent to OpenRouter, which forwards them to the model provider that serves the model we requested. The response gives the probability of a localized AI edit, the coordinates of any suspect regions and the written explanation.
  6. The verdict, confidence, model identifier, explanation, region coordinates and the two Storage URLs are written to your account in our database. The uploaded bytes are then discarded.

The pHash computed in step 2 is stored but is not used to match images to each other. It is deliberately tolerant of small changes, which is exactly the thing this app exists to find — reusing a verdict across two images that differ by a small edit would hide the edit. Only exact SHA-256 matches reuse a result.

When you delete a scan, both compressed copies are removed from Supabase Storage and the database row is removed. When you delete your account, every scan you ever made is deleted along with both compressed copies for each one.

Note on Supabase Storage URLs: the storage bucket is public for performance reasons (your app reads images directly from the CDN). The URLs are not guessable in practice (random UUIDs), and we do not publish them anywhere outside your own app, but anyone you knowingly share a URL with can view the underlying image without authentication. We are evaluating moving images behind a signed-URL gateway in a future update.

5. Data Retention

6. Your Rights

7. What This App Does and Does Not Detect

PixelProbe looks for localized AI edits inside otherwise real photographs — a face swapped, an object added or removed, a background filled in by a generative tool. When it finds one, it marks the region on the image and writes what looks wrong about that area: blending at the edges, blur or sharpness that disagrees with the rest of the frame, lighting and shadows pointing the wrong way, reflections that do not match their surroundings, texture that is too regular to be real.

Two things follow, and we would rather you read them here than discover them later.

"No edits found" is not proof of authenticity. A careful enough edit is invisible to any detector, this one included. A clean result means the model saw nothing it could point at, not that nothing was done.

A flagged region is a suspicion, not a verdict. Heavy compression, aggressive computational photography, screenshots, unusual lighting and low-quality sources can all look like tampering to a model. False positives happen.

PixelProbe is also not built to answer these questions, and its output should not be read as answering them:

Verdicts are probabilistic. Treat a result as one more piece of evidence, never as the sole basis for a legal, journalistic, employment or safety decision.

8. Data Security

All requests use HTTPS. Authentication tokens on iOS are stored in the Keychain with the WHEN_UNLOCKED_THIS_DEVICE_ONLY protection class. Passwords are stored as bcrypt hashes; we never log them, never email them, and never see them in plaintext beyond the in-memory hashing step. Database credentials and third-party API keys live in server-side environment variables only.

9. Children's Privacy

PixelProbe is not directed at children under 13 and we do not knowingly collect data from them. If you believe a child has provided us data, email us and we will delete it.

10. Changes to This Policy

We may update this policy. The current version is always at this URL with a revised effective date at the top. Material changes — new third-party recipients, new data categories, changed retention — will be reflected here before the change takes effect, not after.

11. Contact

Questions, deletion requests, or data-access requests: support@pixelprobe.media

Operator: INAYR LLC, 2730 SW 3rd Ave, Ste 306-34, Miami, FL 33129-2323, United States.