Polymorphic Case study
← All case studies

01 / INSURANCE

Compressing six hours of medical review to 45 seconds.

We built a privacy-preserving pipeline that extracts and validates facts from unstructured medical records, giving underwriters auditable evidence without making the model the decision-maker.

The plateRecord: the system reads a medical file page by page. Facts: every extracted fact includes the line it came from, then gets challenged by an independent pass. Escalation: when confidence is sufficiently low, an underwriter is sent the extracted claim and source for manual review.

ClientLIFE INSURER
CapabilityDOCUMENT INTELLIGENCE
Duration8 WEEKS
StatusDELIVERED

01 / the decision
What information does an underwriter need, and how certain are we?

The bottleneck was reliable interpretation of the records.

Experienced underwriters spent up to six hours on each file, navigating long and inconsistent medical records. Diagnoses, medications, dates, and risk indicators could be buried in scanned letters, tables, abbreviations, or conflicting clinical notes.

They wanted faster turnaround, but the workflow was too important for an opaque summary. Every extracted fact needed provenance, a confidence estimate, and a clear route back to the source.

Fig. 01

Time to process each file

What had been six hours of manual review became under a minute of preparation before the underwriter saw the case.

Lower is better

02 / the build
The pipeline grounds, extracts, challenges, and reviews.

A pipeline with an independent second opinion.

We combined layout-aware parsing with a constrained extraction layer, mapping diagnoses, medications, procedures, dates, and risk flags into a typed schema. Every field kept page-level provenance.

A separate challenger model looked for contradictions, temporal ambiguity, and unsupported inference. We routed low-confidence or disputed fields to an underwriter, along with the relevant source passages.

  1. Stage 0101

    De-identify

    We replaced direct identifiers with reversible encrypted tokens before model processing.

  2. Stage 0202

    Structure

    We parsed scans, tables, clinical notes, and attachments into a source-addressable record.

  3. Stage 0303

    Extract & challenge

    We generated candidate facts, then independently tested them for support, chronology, and contradictions.

  4. Stage 0404

    Escalate

    We sent uncertainty, not volume, to the human reviewer, with evidence already attached.

03 / the result
Measured in time, expert agreement, and privacy

Faster where safe; human where judgement mattered.

To tailor for the desired use case, we separated evidence assembly from underwriting judgement. The system handled document navigation and consistency checks; authorised professionals were left with the final decision and could inspect every claim.

480×

faster evidence preparation

95.3%

agreement with expert review

0

reported PHI exposure incidents in the pilot

We now use the same architecture as a reusable pattern for other high-volume, high-sensitivity records where claims need to remain traceable to their source.