Scan your data with the real engine

Upload a bounded extract, choose the target, and run the LeakLens Python detectors on a configured backend. This page never replays a recorded result and never invents a finding.

live backend required The public site is static. GitHub Pages cannot execute Python. The scan runs only after you connect a real LeakLens API; when none is connected, the page says so and links to the prerecorded demo.

Connection

Point the page at a LeakLens API

Not connected

Use a backend you run locally, or the documented free deployment. The URL is stored in this browser only; no account is created.

No backend connected. Nothing is being scanned.

This GitHub Pages build is static. Run the API locally or deploy the documented service, then connect its URL above.

Open the prerecorded demo Read the API docs Deployment and limits

uvicorn leaklens.public_api:app --port 8000

Dataset

1. Choose the file to audit

validated server-side

Drop a dataset here or browse CSV or JSON by default. The backend validates the content, not the filename.

Held-out test file optional

Add a separate test split to measure exact, feature-vector and near-duplicate overlap, group overlap and temporal boundaries.

Drop the test split or browse Same limits and validation as the training file.

Column preview

Connect a backend to inspect columns and choose a target.

Options

2. Configure and run

Target column Group column optional Time column optional Single-feature threshold 0.95 Scores at or above this value count as target leakage.

Add experimental AI hypotheses optional

Runs the optional offline semantic rules and learned model on the backend. AI output is separate and evidence-bound: it never changes deterministic findings, severity or the trust score. Off by default.

Inspect a file first; target and optional columns are populated from its real headers.

No account. No permanent storage. Findings are returned for this request only.

JavaScript is required for a live scan

Every LeakLens command still works without a browser: leaklens data.csv --target is_fraud. The prerecorded demo also shows real recorded output.

Server limits

.csv .json

4.0 MiB per file · 5,000 rows · 30s scan timeout

Limits are enforced by the backend and can differ from this displayed default after connecting.

What happens to the file

Parsed in memory for one request; never written to disk by LeakLens.
Size, row, column and cell limits are checked before detection.
The backend validates content, not the filename or MIME type.
No account, no tracking, no permanent dataset storage.
Errors are generic; stack traces and internal paths are not returned.

What is checked

Single-feature target leakage.
Identifier-like columns.
Categorical target leakage.
Group leakage risk.
Optional exact, feature-vector and near-duplicate split overlap.
Optional group and temporal boundary overlap.

Honest boundaries

This page does not claim that a scan ran when it did not. If the backend is missing, sleeping, rate-limited or timed out, you get an explicit error and a retry, never a fabricated score.

The GitHub Pages copy is not connected to a hosted backend unless a URL is supplied. The repository includes local and Render deployment instructions.

Compare with the prerecorded demo →

Scan your data with the real engine

Point the page at a LeakLens API

1. Choose the file to audit

Held-out test file optional

Column preview

2. Configure and run

Real scan in progress

Scan failed

Findings from the Python engine

Dataset audit

Split impact

AI intelligence experimental

Findings