filehugger.comSensitive Data Finder
Find candidate email addresses, phone numbers, IBANs, payment-card numbers, Turkish identity numbers, IPs and URLs before sharing.
Analyse your own file
Choose a PDF, text/data files and images with optional OCR file
Drop it here or browse this device
Your original file is analysed in this browser and is not uploaded. An optional AI second-opinion button appears only after the local report; if you choose it, extracted text—not the original file—is sent to our server.
What happens to the file
- Default processing
- Browser-only JavaScript, PDF.js, pdf-lib and Tesseract where the task needs them
- Accepted input
- PDF, text/data files and images with optional OCR
- Report contract
- File Intelligence JSON schema 1.0 plus a flattened CSV view
- Optional AI
- Explicit second opinion; extracted text only; original file is not uploaded
Synthetic data contains one email, one Turkish IBAN-like value and a clearly marked test payment-card number. No real person’s data is used.
How the result is calculated
Text is extracted locally, then bounded regular expressions and checksum rules find candidates. Payment cards use Luhn; Turkish identity candidates use both checksum rules. Values are masked in exported findings.
Confidence combines pattern specificity and checksum validation. A match is a review candidate, not proof of identity or sensitivity. Risk score weights payment cards highest, then checksum-valid IDs/IBANs, then contact and network data.
Supported and unsupported cases
| Input or situation | What version 1.0.0 checks | What it does not claim |
|---|---|---|
| Selectable PDF/text | Emails, phones, IBAN-like strings, Luhn-valid cards, checksum-valid Turkish IDs, IPs and URLs | Names and addresses without explicit patterns |
| Scanned image/PDF | Same patterns after optional local OCR | Reliable handwriting recognition |
| Structured data | Visible scalar text | Hidden database relationships or external references |
| AI second opinion | Optional category review of extracted text | Guaranteed detection or legal classification |
What the confidence score means
Confidence describes how much of the file the published parser could inspect and how specific the evidence was. It is deliberately separate from the result score. A 95/100 confidence value does not mean “95% likely to be true,” and a high health/readiness score does not certify safety, compliance or factual correctness.
Comparison with other approaches
| Approach | Useful strength | Important limitation |
|---|---|---|
| Deterministic local scan | Local, explainable offsets/checksums, masked exports | Limited semantic understanding |
| AI-only review | Can understand context | Non-deterministic and may miss exact strings |
| DLP enterprise system | Policy, central audit and broad integrations | Cost and administrative setup |
| Manual search | Good for known terms | Easy to miss variants and checksum-valid numbers |
7/7 fixtures passed (100%)
Dataset: FH-INT-SYNTH-2026-07. Scope: Email, phone, IBAN, Luhn card, Turkish-ID checksum, IP and negative control. Fixtures contain generated data only; no visitor file or personal record is collected.
When this result can be wrong
Order numbers can resemble phone numbers, valid-looking test cards match intentionally, names have no universal pattern, and OCR can create or hide matches. Always review the source and use true redaction rather than drawing a black rectangle.
Keep the original file, test the output in the receiving application and escalate high-stakes legal, financial, medical, accessibility or security decisions to a qualified reviewer.
Version and accuracy change log
| Release | Date | Calculation/coverage change | Re-test requirement |
|---|---|---|---|
| 1.0.0 | 2026-07-29 | Initial public engine, report schema, thresholds, synthetic fixtures and sample files. | All published fixtures and one browser sample per tool. |
Downloadable evidence, not a black-box verdict
After analysis, use Download JSON report for structured evidence or Download CSV report for review in a spreadsheet. Both include engine version, timestamp, score, confidence, metrics, findings, recommendations and limitations. Sensitive-data reports mask candidate values and no report includes the complete extracted document text.
Related tools and independent checks
Method or accuracy correction? Read the testing methodology and contact info@brothers.net.tr. Material scoring changes are published in the change log.