Claims Intelligence
A photo, a voice note, a half-filled form — compact on-device specialists turn messy claim inputs into a filing-ready FNOL packet with zero-hallucination coaching.
Claims Edge SDK
Intelligent intake inside any claims surface
Liquid is the on-device claims intelligence layer that drops into carrier portals, partner mobility apps, travel portals, and health reimbursement flows. The host app owns the claimant journey; the SDK turns photos, documents, and voice into structured packet data, flags missing evidence before submission, and emits a clean handoff for examiners.
Deployment envelope
iOS SDK- Shared runtime: 219.3 MB
- Vision projector: 102.8 MB
- FNOL vision extractor: 26.1 MB
- Transcript extractor: 28.7 MB
- Coach classifier: 28.9 MB
- Audio ASR uses native iOS Speech
- Lineage verifier passes for canonical LFM2.5-VL-450M
Two examples
From SDK architecture to working modules
Example 01 — FNOL evidence extraction
Photos, documents, and a voice note become structured fields with confidence, source provenance, and an adjuster-ready packet.
Open FNOL demoExample 02 — Claimant coaching before submission
A compact classifier detects delay and denial risks, then uses carrier-approved catalog language to tell the claimant what to fix.
Open coach demoThe clean-claim path is mostly solved. What carriers struggle with is the gray zone: eligibility uncertainty, missing evidence, and denial-driving ambiguity. A compact claims SDK combines vision extraction, transcript field extraction, and a fine-tuned coach classifier to structure photos, documents, and voice notes, then surface coverage gaps with coaching text from a locked legal-vetted catalog. No LLM-generated language reaches the policyholder. PII never leaves the device.
2 specialist models
How It Works
One demo proves multimodal FNOL extraction.
One demo isolates the claimant coach before submission.
Photos, Voice, Documents — Structured Fields in Seconds
FNOL is the moment a claim is born — a damage photo, a voice note, a half-filled form. Adjusters spend 15-20 minutes per call gathering information that's already in the evidence. A 450M vision model extracts vehicle registrations, damage severity, and policy numbers from uploaded photos. A 1.5B audio model transcribes incident narratives into structured fields. What took manual data entry now happens in seconds, with field-level provenance and a complete audit trail.
The Gray Zone — Denial Risks That Rule Checklists Miss
The clean-claim path is solved. What carriers struggle with is eligibility uncertainty, missing evidence, and denial-driving ambiguity that manual checklists can't catch. A fine-tuned 350M classifier identifies cataloged gap types — ambiguous onset dates, untranslated invoices, missing third-party insurer details — before the claimant submits. The model predicts the gap ID; carrier-approved catalog language tells the claimant what to attach, clarify, or correct. Zero hallucination. No adjudication language.
Try each model
All Demos
Claims FNOL
A photo, a voice note, a half-filled form — three on-device models turn messy claim inputs into a filing-ready FNOL packet with denial-risk coaching.
Zero hallucination by construction — classifier predicts gap IDs, catalog provides every word the policyholder sees.
Claimant Coach
A coach-first FNOL experience that shows how an on-device model helps claimants fix incomplete packets before submission.
The model identifies the gap; the legal-vetted catalog speaks to the claimant.
Ready to deploy in your environment?