On-prem CPU / GPU
Full throughput on commodity servers — encoder models run in milliseconds on CPU, faster still with a single GPU.
Open-source insurance NER
Your Data. Your Model. Your Hardware.
Extract every policy, claim & party entity — and strip GLBA + HIPAA data — 100% on your hardware.
OpenCover is a family of eight trained encoder models built for insurance. They read dec pages, FNOL notices, ACORD forms, adjuster notes and medical reports, extract 28+ entity types, and de-identify PII and PHI without a single record leaving your network. Apache-2.0, hosted on Hugging Face.
8
insurance models
all trained
GLBA + HIPAA
dual-regulation
de-identification
~150 ms
per document
on-device
$0 · 0
API cost · records leave your network
Quickstart
Four lines from raw document to safe, structured data.
- One import, one call. Pick a compliance profile and OpenCover extracts, labels and redacts in a single pass — entirely in your process.
$ pip install git+https://github.com/flowx-ai/openner
from openner import analyze result = analyze(document, profile="glba-hipaa") print(result.entities) # POLICY_NUMBER, CLAIM_NUMBER, PERIL, VIN … safe = result.deidentify() # strips PII + all 18 HIPAA identifiers on-device
Runtime
Runs everywhere your documents already live.
Small encoders mean no GPU cluster, no vendor API, no data-processing addendum. Deploy at the point of intake.
Air-gapped
Ship the weights on physical media and run with zero outbound connectivity — built for regulated core networks.
Inside your claims / policy admin system
Embed as a library in Guidewire-, Duck Creek- or homegrown-core pipelines, right where FNOLs and dec pages arrive.
Browser (ONNX / WASM)
Quantized ONNX builds run in-browser via WASM — redact a document before it ever reaches your server.
REST / gRPC
A one-command self-hosted service exposes every model behind REST and gRPC for your internal platform teams.
Privacy & de-identification
Dual-regulated data never leaves your premises.
Insurance is the rare industry that handles financial NPI and health PHI in the same claim file. OpenCover treats both as first-class.
One model, both regimesInsurRedact covers GLBA financial PII and all 18 HIPAA Safe Harbor identifiers in a single pass — no chaining two redaction vendors.
Insurance-shaped identifiersStrips SSNs, driver's license numbers, VINs, bank and payment details, plus diagnoses and ICD/CPT codes buried in medical-claims narratives.
Configurable policy profilesShip-with profiles for glba-npi, hipaa-safe-harbor and gdpr-art9 — or compose your own per line of business.
Proven clinical lineageThe medical-claims branch reuses battle-tested clinical de-identification weights, adapted to bodily-injury and health-claim language.
Signed audit reportsEvery de-identification run emits a signed, replayable report — evidence for Fair Claims Practices reviews and DOI market-conduct exams.
Zero data movementNo cloud calls, no telemetry, no usage phone-home. The only network traffic is the one-time model download — and even that is optional.
Model library
Small models for insurance, one job each.
All eight models are trained and shipping — compact encoders with ONNX/edge exports and built-in VIN check-digit and ICD/CPT validation. Download from Hugging Face once, fine-tune freely, compose as needed. For scanned or photographed forms, the cross-industry FlowX DocFormNER model is an early preview (see the FAQ).
PolicyDetect
Policy numbers, coverages, limits, deductibles, premiums and effective dates from declarations pages.
POLICY_NOLIMITDEDUCTIBLEPREMIUM
ClaimExtract
Claim number, loss date, cause, reserve and status from FNOL notices and running claim files.
CLAIM_NOLOSS_DATERESERVE
PerilClassify
Tags perils and cause-of-loss — fire, flood, wind, theft, collision — from free-text loss descriptions.
PERILCAUSE_OF_LOSS
PartyResolve
Roles every party in the file: policyholder, claimant, adjuster, broker, beneficiary, carrier.
POLICYHOLDERCLAIMANTADJUSTER
AssetDetect
Insured objects and their identifiers: VIN, make/model/year, property addresses, equipment serials.
VINPROPERTY_ADDRSERIAL
CodeMap
Reference codes wherever they hide: NAIC, NCCI class, ICD/CPT, CAT event, NAICS/SIC.
NAICICDCPTCAT_CODE
InsurRedact
The dual-regulation de-identifier: GLBA financial PII plus all 18 HIPAA Safe Harbor identifiers.
SSNVINDLICDPHI
MedClaimNER
Clinical branch for bodily-injury and health claims — reuses proven clinical NER weights.
DIAGNOSISICDPROVIDERINJURY
Built with OpenCover
From toolkit to product in one sprint.
Teams compose the models into working products. Three we ship as reference implementations:
FNOL Intake Copilot
Turns a phone transcript, email or photo FNOL into a structured, ACORD-mapped claim in seconds — every PII field flagged before it hits the queue. Runs on-prem.
Claims-Triage Agent
Reads adjuster notes and medical reports, scores severity, and routes fast-track versus SIU/fraud review — with a full audit trail for Fair Claims Practices.
Underwriting Submission Reader
Ingests broker slips, loss runs and SOVs; extracts risk entities and codes into a clean submission summary — no data ever leaving the carrier.
Compliance
Designed for the rules insurance actually answers to.
On-device processing doesn't exempt you from regulation — it just makes compliance tractable. OpenCover maps to the frameworks your legal team will ask about.
HIPAASafe Harbor de-identification of health-claim PHIGLBA + NAIC #672protection of nonpublic personal financial informationNAIC Model Laws#668 Insurance Data Security Model Law alignmentState DOI rulesdata-handling requirements across state insurance departmentsFair Claims Practicesauditable, replayable claim-handling decisionsGDPRArticle 9 special-category health data, processed locallyEU AI Acttransparency for high-risk underwriting and claims AISolvency II & IFRS 17auditable data lineage from document to ledgerACORDextraction targets map to the ACORD data model
Benchmarks
Measured against frontier LLMs.
We benchmarked our small on-device models head-to-head against the latest cost-effective frontier LLMs — same documents, same entity types.
6–24×
faster than frontier APIs
$0 vs $0.03–5.57
per 1k documents
0
records leave your network
Frontier LLMs win at open-ended reasoning — not structured, convention-bound extraction. Small fine-tuned encoders win there, and keep dual-regulated data on-prem.
FAQ
Questions carriers ask first.
Q.01What is OpenCover?
An open-source, privacy-first family of eight trained encoder NER models for insurance — PolicyDetect, ClaimExtract, AssetDetect, PerilClassify, PartyResolve, CodeMap, InsurRedact and MedClaimNER — covering P&C, life, claims and underwriting. They extract 28+ entity types with VIN check-digit and ICD/CPT validation, and de-identify GLBA + HIPAA data with signed audit reports, all on your own hardware. ONNX/edge exports included. Apache-2.0 licensed, models hosted on Hugging Face.
Q.02Why encoder models instead of a large LLM?
Extraction and redaction are token-classification problems, and small encoders are excellent at them: deterministic spans, millisecond CPU inference, no hallucinated fields, and models small enough to run inside a claims system or even a browser. Use an LLM downstream if you want; OpenCover keeps the sensitive first pass local and cheap.
Q.03Why dual GLBA + HIPAA de-identification?
Insurance sits at the intersection of two regulatory regimes: policy and payment data is financial NPI under GLBA, while bodily-injury and health claims carry PHI under HIPAA. A single claim file often contains both. InsurRedact handles them in one pass so you don't have to stitch a banking redactor onto a healthcare one.
Q.04Does any record ever leave my network?
No. Inference runs entirely in your process on your hardware. There are no cloud APIs, no telemetry and no phone-home. The only network activity is the one-time model download from Hugging Face — and in air-gapped mode you skip even that by importing the weights on physical media.
Q.05Can I fine-tune on my own claims data?
Yes — that's the point of shipping small models under Apache-2.0. Fine-tune on your carrier-specific forms, endorsements and note-taking conventions with standard Hugging Face tooling, entirely inside your network. Your data never leaves; your improved weights are yours.
Q.06Can it read scanned or photographed documents?
As an early preview, the cross-industry FlowX DocFormNER model (a LayoutLMv3 encoder that reads text, 2D layout and the page image together, published at huggingface.co/flowxai/docformner) now reads scanned and photographed insurance documents such as ACORD forms and FNOL/claims packets, entirely on-prem — though it is trained on synthetically rendered forms and still needs evaluation on your real OCR'd scans before production use.
Q.07How are the models trained, and is it really open source?
Everything — code, weights, training recipes — is Apache-2.0. An honest caveat: public insurance-text NER corpora are scarce, so training relies heavily on synthetic generation of ACORD forms, FNOL notices and dec pages, plus reuse of clinical de-identification corpora (i2b2/n2c2) for the medical-claims branch. We publish the generation pipelines so you can inspect and extend them.
Apache-2.0 · An open-source model family — not affiliated with any carrier or regulator.