Ingestion Agent
standing by…
not reached
Agent demo · fictional tenant
Multi-agent claim validation
Kestrel Insurance sends claims to a portal. Four agents take it from there — supervised, budgeted, and traceable.
new
›
ingested
›
extracted
›
validating
›
decided
Email
Chat
File drop
POST /docs
API
Dify
RPA / workflow
Claims portal
no API? no problem ↓
Intake queue
idle · 4 waiting, deduped
Orchestrator Agent
standing by — pick a scenario and run the pipeline
›
Extraction Agent
standing by…
not reached
›
Validation Agent
standing by…
not reached
›
Decision
—
FRAUD FLAG
awaiting run
Email
stand by
Chat
stand by
File drop
stand by
webhook
API
stand by
Dify
RPA / workflow
stand by
Claims portal
stand by
Kestrel claims portal
Step 2 of 3 — attach supporting documents
Claim reference
CLM-2041
Invoice document
hospital-invoice.pdf
Submit claim
recording operator session · 12 actions
CITYPHYSIO#R-8841
Patient: J. Fischer
Physiotherapy — 4 sessions
Service date: 2026-04-10
AMOUNT $310.00
0.96
SIMILARITY
threshold 0.90
comparing…
CITYPHYSIO#R-8841
Patient: J. Fischer
Physiotherapy — 4 sessions
Service date: 2026-04-10
AMOUNT $310.00
Covers
Step 2–4 · Extraction → Decision
not run yet
for the claims reviewer
What the pipeline reads off the document, and what it decides.
Pick a scenario and run the pipeline — the decision lands here.
FRAUD FLAG
confidence
Reasoning
rule_iddescriptionparameterstatus
44
R-201
Claim submitted within 90 days of treatment (fast-track)
window: 90d
—
45
R-203
Claimant identity matches policy holder
exact match
—
46
R-205
Claim currency is supported
USD · EUR · JPY
—
47
R-207
Claims must be submitted within 90 days of treatment
window: 90d
—
48
R-210
Claim ID unique within tenant
24-month scope
—
49
R-214
All referenced attachments present
required docs
—
Rule sets are ingested directly from each tenant’s Excel files. A few hundred rules per tenant is typical — Kestrel Insurance runs 214 across 4 sheets.
Run the pipeline to capture an execution trace.
—
straight-through processing
—
avg turnaround
—
claims this week
—
in review queue
Live claim inbox
streaming
Last 7 days
processed
escalated
All three doors — live at once · people in the UI, machines over API and MCP
web UI reviewer opens the run, approves the claim REST POST /v1/claims/validate → 202 CLM-2041 MCP tools/call validate_claim(CLM-2041) three doors, one engine — same rules, one audit trail
Drops into the automation stack you already run
Sign in with SSO
Your identity provider, your roles.
Work the queue
Only what the agent escalated.
Nothing to integrate
AuroraAgent is the app — day one, no build.
n8n
Node in a workflow.
UiPath / RPA
Legacy desktop apps with no API at all.
Dify
Tool inside an agent app.
Power Automate
Outlook and SharePoint intake flows.
Inbound tools
upload_documentPush a claim document into the pipeline for a tenant.
get_claim_statusRead the live state machine position of any claim.
validate_claimRun the rule set against an extracted claim.
approve_claimHuman-in-the-loop approval, recorded with the reviewer id.
reject_claimReject with mandatory rule citations attached.
Outbound tools
push_resultDecision + citations back to the tenant portal.
notify_staffReview-queue and decision notifications to claims staff.
export_auditFull trace to the tenant’s system of record.
Extracted document
hover a row or a box · click to zoom
FieldValueConf
Review cockpit
NEEDS REVIEW
Receipt — Physio & Wellness
Patient: K. Okafor
8 sessions physiotherapy, lower back
May–June 2026, paid cash
Total: $1,240amount · 0.62
thank you!
Fields · 3
amount
Enter ⏎
Corrections are captured and fed back for model improvement.