De-identification

Your company licenses a copy of its entire operating record: mail, chat, files, tickets, and ledgers. Mikra transforms all of it so that identities change and the work stays intact. One real person becomes one stand-in identity in every source. You review a sample before any customer receives anything. The risk is lower, not zero.

Mikra / De-identification · SampleIllustrative
Email · Mar 14, 2024
Before: From: Dana Whitfield <dana.whitfield@harborline-cpa.com>. Sending the bank reconciliation to Kessler Dental today.
After: From: Priya Raman <priya.raman@alderbrook.example>. Sending the bank reconciliation to Marlow Dental today.
Chat · #close-team
Before: @dana.w Kessler Dental says invoice 4471 was paid twice.
After: @priya.r Marlow Dental says invoice 4471 was paid twice.
Ticket · AP-2291
Before: Assignee: Dana Whitfield. Customer: Kessler Dental. Call (617) 555-0148.
After: Assignee: Priya Raman. Customer: Marlow Dental. Call (000) 555-0199.
Same person, 3 sources
One stand-in: Priya Raman
Same client, 3 sources
One stand-in: Marlow Dental
Raw upload
Deleted within 120 days
Illustrative sample / Every name and record is invented
01 / The problemWhy finding names is not enough

Black bars break the record.

Business records are valuable because they hold long workflows and relationships: who asked, who approved, who was late. They also hold personal information. Blacking it out lowers the risk, but it strips context and breaks the links between records. So the job is not finding names. It is transforming a whole body of records, with three goals.

02 / Five approachesCompared on three things

Five methods, compared.

There are five common ways to handle personal information in records. Each trades privacy against how useful the result is and whether it still holds together across documents.

Five ways to handle personal information, rated low, medium, or high on privacy, usefulness of what is left, and consistency across records
ApproachPrivacyUsefulness of what is leftConsistency across records
Removal or blacking outThe value is deleted or covered.HighLowLow
Pattern maskingAnything shaped like an email or a number is replaced with a fixed mask.MediumLowLow
Entity replacement, field by fieldEach detected value is swapped on its own.Medium to highMediumMedium
Fresh synthetic dataNew records are generated to resemble the originals.HighMediumLow
Identity-preserving transformationEach person or company gets one stand-in identity across the whole record.Mikra’s approachHighHighHigh

Only the last approach protects identities and keeps the record usable as a whole. The ratings describe each approach done well. They are not measurements of any one system.

03 / How the transformation runsMikra’s method, stage by stage

Seven stages, in order.

This is the process Mikra runs on a company’s records. Machines do the first pass. People check the hard cases and a sample of the result. You check last.

  1. 01

    Inventory and segmentation

    Every file is listed with its format and what it takes to rebuild it. Large files are split into parts. The inventory is the count everything is measured against later, so a file that fails cannot quietly drop out.

  2. 02

    Detection and entity resolution

    Detectors flag candidate identifiers across the whole body of records, not one file at a time. A second sweep links aliases and indirect mentions to the same person or company. That linking is called entity resolution. Finding a value and deciding whose it is are separate steps, tied together by one shared identity record.

  3. 03

    Review of hard cases

    Ambiguous candidates are confirmed, rejected, or relabeled. Aliases are merged. False merges are split. Degraded values, such as typos, look-alike characters, and odd spacing, are traced back to their owner.

  4. 04

    Identity mapping

    Each protected person or company gets one stand-in identity: a name, an email address, a username, and an employer that fit together. It is not a set of random field-by-field swaps. The mapping is fixed for the whole run.

  5. 05

    Reconstruction and validation

    Files are reassembled and checked against the inventory: rows and columns, sheet names, page counts, archive members. A file that opens is not proof that the identity inside it is right. So structure checks and meaning checks are separate.

  6. 06

    Human review

    Reviewers read a sample drawn from the inputs, not from the successes, so failed or missing files stay in the measurement. They mark exposed identifiers, unnecessary changes, wrong labels, fragmented replacements, two people merged into one, and anything that reads wrong.

  7. 07

    Your review

    Before the first delivery to any customer, you get a representative sample of the transformed data and a short description of the methods. You have 10 business days to point out anything we missed. We fix it before anything is delivered.

04 / Who touches whatStandard track

Raw data stops at Mikra.

Three parties, in order. Your company and Mikra hold the raw records. Customers receive only the de-identified data and what is built from it.

Custody flow: Company systems (Mail, chat, files, ledgers), then Secure upload (Sent to Mikra, encrypted), then Mikra (Transforms identities), then Your review (Sample, 10 business days), then Customers (De-identified data only). Raw data stops at Mikra, which transforms it on its own Ceph storage cluster in a colocation data center. Your sample review and Mikra’s customers see only de-identified data.
05 / How we judge the resultFive questions

Five questions, answered separately.

A strong score on one question cannot hide a failure on another, so the five are reported separately. Mikra will report them to your company with its sample. We have not yet published measurements of our own, so no score appears here.

06 / The evidenceWhy the risk is lower, not zero

Detection alone falls short.

These charts are about the problem, not about Mikra. They show how much a first pass misses, and how much entity resolution and review recover.

Source of every figure in this sectionPublished third-party evaluation, 2026. None of these systems is Mikra’s. Mikra has not yet published its own measurements.

Chart A / Third party, 2026Not Mikra’s results

Off-the-shelf detection still misses identifiers.

Seven systems ran on a public test set of 5,198 workplace messages holding 8,805 hand-labeled identifiers of five types: given name, family name, email address, username, and organization. Two measures are shown. Recall is the share of known identifiers a system found. Precision is the share of its flags that were correct.

Off-the-shelf detection still misses identifiers. Published third-party evaluation, 2026. None of these systems is Mikra’s.
SystemIdentifiers found (recall)Flags that were correct (precision)
Purpose-built transformation pipelineBest in the evaluation95.1%97.0%
Commercial de-identification tool92.6%96.5%
Open-source rule-based detector89.0%87.3%
Open-source entity-recognition model88.6%85.2%
Large general-purpose AI model88.2%95.4%
Mid-size general-purpose AI model85.0%97.0%
Small general-purpose AI model79.9%97.6%
The best system in that evaluation still missed about 5 in 100 identifiers on a first pass, and precise but cautious tools missed up to 20 in 100. That gap is why detection is only the first stage, and why review, your own sample, and the contractual bar on re-identification exist.
Chart B / Third party, 2026Not Mikra’s results

Messy enterprise data is harder.

A synthetic company dataset was built to be hostile: 600 employees and 420 customers recurring across HR, finance, technology, logs, and customer tables. Identifiers were hidden behind typos, look-alike characters, spelled-out digits, odd separators, and zero-width characters, which take up no space on screen. Decoys were planted, such as part numbers shaped like Social Security numbers and version numbers shaped like IP addresses. A widely used open-source detection tool ran in its default configuration, then again with entity resolution across the whole dataset and a review step.

Messy enterprise data is harder. Published third-party evaluation, 2026. None of these systems is Mikra’s.
MeasureDefault configurationWith entity resolution and review
Planted identifiers foundHigher is better74.3%89.2%
Flags that were correctHigher is better62.4%77.5%
Look-alike decoys wrongly transformedLower is better72.3%40.0%
Entity resolution and review measurably help. They do not make the result perfect.
One hard case

Some cases cannot be settled.

Real data wears down. A title such as “Ms.” can turn up as “MS.”, “M.s”, or just “M.”. So can “Mx.”. Once both have collapsed to “M.”, the information that told them apart is gone. Link the two records and you may merge two people into one. Keep them apart and you may split one person in two. No amount of care recovers what the data no longer holds.

Ms.
MS. · M.s · M.
Mx.
M.
08 / What is preservedThe structure a model learns from

The work stays whole.

A model team trains on how work moves from request to approval. Strip the structure and nothing is left to learn.

09 / What is never uploadedExcluded records

Some things never leave.

These records are never uploaded to Mikra. Patient or clinical records and a law firm’s client files move only through the healthcare and legal track.

10 / Residual riskStated plainly

The risk is lower, not zero.

It is lower for three reasons. What identifies a person or your company is transformed: names, company and client names, contact details, account numbers, and similar identifiers. You review a sample before anything is delivered. Customers are contractually barred from re-identifying anyone. No method finds everything, and we put no number on our own work until we have measured it.

Data partnerships

Your records,
your review.

Tell us what you hold. We assess it and walk you through this process before you sign anything.