Unstructured Data Mining
Finding the hidden tortfeasor inside unstructured claim text.
Adjusters write down the third party. They just write it on page three, in shorthand, at 6pm. Ishqkala reads all of it.
Read in the last 24 hours · one tenant
- Diary entries
- 184,220
- Police reports
- 6,914
- O&C / engineering
- 312
- Estimates
- 21,775
Median time from document arrival to scored finding: 4 minutes.
See the extraction
Every finding points back to its source span.
CLM-25-077310 · Commercial property · diary entries 4 to 6
04/02 Insd mgr states HVAC contractor replaced condensate pump on 03/28. Overflow switch not wired to shut off unit. Water through ceiling tiles into server room. Contractor's service ticket #4471 in file. BI claim pending, 3 days downtime.
Extracted · linked to source span
- PartyHVAC service contractorWork performed 5 days before loss
- Liability signalOverflow safety switch not wiredInstallation below code (IMC 307.2.3)
- EvidenceService ticket #4471Identifies contractor and work date
Data sources analyzed
Seven document families, one claim picture.
- Adjuster diary entries and supervisor notesFree-text diaries with claims shorthand (IV, OV, CV, DOL, NOL, PD, BI) resolved to parties, dates and events.
- Police collision and accident reportsNarratives, diagrams, citations and contributing-factor fields, including scanned and handwritten pages.
- Fire origin and cause (O&C) reportsArea of origin, ignition source, arc mapping findings and artifact retention notes from NFPA 921 investigations.
- Structural and mechanical engineering reportsForensic findings on failed components, installation defects, code violations and workmanship.
- ISO ClaimSearch and adverse carrier matchesPrior claims, other-party policy information and carrier identification already delivered to your claim file.
- CPSC and NHTSA safety recall databasesModel, serial and VIN details cross-matched against product and vehicle recall notices.
How the parser works
Five passes over every document.
- 01
OCR & layout
Pages, tables, diagrams and handwriting become structured text with confidence scores.
- 02
Claims vocabulary
Shorthand and carrier-specific abbreviations resolve to parties, vehicles and events.
- 03
Party & role extraction
Contractors, drivers, owners, employers, manufacturers and public entities, each with a role.
- 04
Liability signals
Admissions, citations, recent work, failed components and code violations are classified.
- 05
Evidence pointers
Each signal links to its page and span, with custody status for physical evidence.
- 80%+
- of AI-found opportunities come from unstructured narrative text, not structured fields
- Ishqkala detections
- 100 to 150
- active claims per P&C adjuster
- Carrier Claims Workload Study
- 4 min
- median document-to-finding latency on open claims
- 92%
- adverse liability scoring accuracy, verified by subrogation counsel
“The tortfeasor was in the officer's narrative the whole time. Our keyword rules never fired because nobody types 'third party' in a crash report.”
FAQ
Frequently asked questions
No. Most recoverable third parties never reach a structured field. Ishqkala reads the narrative and documents directly, and over 80% of the opportunities it identifies come from unstructured text rather than claim system fields.
Every page goes through OCR and layout analysis. Handwritten spans carry a per-word confidence score; anything below threshold is flagged for a quick human check instead of being guessed.
Yes. Every extracted party, liability signal and evidence item links back to the exact document, page and span it came from, so a subrogation specialist can verify a finding in seconds.
No. Customer claim files and adjuster notes are never used to train any model. Processing runs inside your isolated tenant in US AWS regions.
Find the millions hiding in your closed claim files before the Statute of Limitations expires.
We'll scan a batch of 500 closed claims for free and deliver an itemized recovery manifest within 5 business days.