Detection
Did it find everything
Precision and recall per entity type. Model quality never reaches perfection, and being honest about that is why the human checkpoint exists.
Use case
Detection proposes. A person decides.
Case files, personnel records and legal disclosure leave the building only with personal data removed. Detection proposes; a person decides; approved regions are deleted.
Four stages, one checkpoint that matters.
Everything before the human speeds the review up; everything after it is mechanical and verifiable.
01
Detect
PII found across every page, scans included, with page and region evidence.
02
Group
Repeated values collapse into one decision instead of fifty approvals.
03
Review
A person keeps or removes each finding against its evidence.
04
Apply
Approved regions are deleted from a copy; reopen the output and check.
Two things worth measuring separately.
Detection quality and removal completeness are different risks with different properties. Collapsing them into one figure hides whichever is weaker.
Removal
Was the approved region actually removed
Once a reviewer approves a region, its content is gone from the file, not covered by a rectangle a copy-paste defeats.
Why it matters
One number hides the risk
A single accuracy figure lets a strong removal guarantee mask weak detection, or the reverse. Disclosure risk needs both, separately.
Where the human actually belongs.
The reviewer is the point of the workflow, so the workflow is built to spend their attention well.
Grouping
Repeated values, decided once
The same name appearing forty times is one decision, not forty. Findings group by value into a short working list.
Evidence
Shown in place on the page
Each finding is shown on the source page with its context, so keep-or-remove is judged with the document in view.
Decisions
Keep, remove, or record why
A reviewer can keep a finding and record why. Values that cannot be placed on a page surface for manual handling.
An artifact you can verify.
Removal, unlike detection, is checkable on the output. So it is checked.
Permanent
Content removed, not covered
Approved regions are deleted from the file. Reopen the output, search it, and the text is not there.
Original
Preserved, untouched
The source file is preserved, so redaction produces a derived artifact, which is what an audit trail needs.
Verified
Checked after the write
The output is re-examined after redaction, so a failure to remove is caught before the file is handed on.
Approved findings, one call, a report.
Reviewer-approved findings become search terms or explicit regions; one call removes them from the file and reports every change.
using LMKit.Document.Pdf;
// Findings the reviewer approved for removal.
var request = new PdfRedactionRequest();
request.SearchTerms.Add("DE89 3704 0044 0532 0130 00");
request.SearchTerms.Add("+49 170 555 0143");
var report = PdfRedactor.RedactToFile(
"claim-file-0472.pdf",
"claim-file-0472.redacted.pdf",
request);
// The report lists every removed region, page by page.# Detect on LM-Kit One, then apply the approved removals.
curl http://your-server:5189/lmkit/v1/detect-pii \
-H "Authorization: Bearer your-server-token" \
-H "Content-Type: application/json" \
-d '{"input":"<document text>"}'
curl http://your-server:5189/lmkit/v1/pdf-redact \
-H "Authorization: Bearer your-server-token" \
-F "file=@claim-file-0472.pdf"The technology underneath.
The workflow composes documented capabilities you can inspect on their own.
Detect
PII extraction
Typed entities with value, occurrences and confidence, page by page.
How PII detection worksReview
Smart redaction
Findings become reviewable regions, grouped by value, decided once.
How smart redaction worksRemove
PDF redaction
Approved regions deleted from the file: text, images, and vectors alike.
How removal worksThen automate
From workflow to watcher
The same detection runs continuously on intake, drafting redactions for review the moment documents arrive instead of the week before disclosure.
The monitoring use caseFrequently asked questions.
Is the redaction really permanent?
Yes. Approved regions are deleted from the output file rather than covered: text glyphs, image pixels, and vector content are removed, and the output is re-examined after the write so a failed removal is caught before the file is handed on.
Why is a human reviewer part of the workflow?
Because detection quality and removal completeness are different risks. Detection never reaches perfection, so a person decides each grouped finding against its evidence; removal, which is checkable, is then verified mechanically.
Does it work on scanned documents?
Yes. OCR runs before detection, so PII in scans and photographs is found, shown in place on the page, and removable like any other region.
Can we run this as a service for several teams?
Yes. LM-Kit One serves detection and redaction as authenticated REST endpoints with audit, so case-handling systems across departments share one governed workflow; LM-Kit.NET embeds the same calls when the workflow lives in your own application.
Private Document Intelligence