BUILD RECORD

PagePurger

Medical-legal review is billed by the page. The money is in the pages nobody has to read. A specialized pipeline that flags and extracts only the non-routine pages in massive discovery.

OVERVIEW

product build   document pipelines   medical-legal review   batch processing   retired

Fewer pages to review, and the reason each removed page came out.

PagePurger

Retired

What it is

A production pipeline for medical-legal record review. Every page got one relevance call, and fingerprints caught the duplicates.

Impact

Every duplicate and unrelated page pulled out was $3 nobody billed. Each came back watermarked, its reason on a spreadsheet row.

Scale

8,000+ pages in one case file, read by hand at about $3 each.

Stack

aws batch gpu   bedrock   perceptual hashing   sftp

My role

Architecture, pipeline, classification, multi-tenant platform, operations.

Verified

Each tenant ran separately; review sets went back over SFTP.

THE METHOD

How a case file became a review set

It pulled blank separator sheets, the same fax scanned four times, and rescans a hair off.

A CASE FILE · THE PATH TO REVIEW
ARRIVED · JUDGED · COLLAPSED · RETURNED

01

Took the case file

Records arrived over SFTP into the tenant's pipeline.

02

Judged every page against the case

Each page got one relevance call, Claude Haiku on Bedrock.

03

Collapsed the duplicates

A file fingerprint found exact copies. Perceptual hashing found near-identical rescans.

04

Handed back the review set

It returned review pages, watermarked exclusions, the decision spreadsheet, and the invoice.

LIMITS STAY ATTACHED

What it refused to do

Nothing was deleted. Every excluded page came back watermarked with its reason, for the lawyer who needs it later.

A long job checkpointed and resumed where it stopped.

Tenants stayed separate. No case or client appears here.

RECORD

Where it went

Status

Retired. Built and run solo.

Sibling build

aiPDF Engine, the same pipeline without the case rules.

NEXT

Back to Side Projects → Read the aiPDF Engine build →

THE ARCHITECTURE

Architecture map of PagePurger: the real components and how they connect
THE ARCHITECTURE MAP

A RELATED PROBLEM

Is page-by-page review setting the cost of the work?

This build is retired, but the review problem remains. If a system can remove obvious waste and show the reason for every removal, it is worth a conversation.

Email me →