Architecture

Follow one claim through the system.

Forty-two thousand claims a month. Fraud screening reaches maybe 5% of them — that is all the hours the adjusters have.

Six steps, in order. Each layer reports what it actually did with the claim.

press play

3.4 s End-to-end — CLM-2041 arriving to decision posted back

11 spans Every tool call, token cost, and rule traced and replayable

Zero egress Nothing leaves your VPC except allowlisted portal traffic

Kestrel’s account, Kestrel’s keys. Eastgate operates the stack, but the data never changes custody.

00 Kestrel Insurance

Policy & claims system

Mailboxes, Slack, file shares

Provider portals (no API)

Claims reviewers

CLM-2041 submitted by the provider — 3 scanned files, 2.4 MB

AuroraAgent — Kestrel’s VPC · ap-northeast-1

01 per tenant

Tenant surfaces

The screens your submitters and your reviewers actually touch.

Claims portal submitters Staff console reviewers Review queue human-in-the-loop Audit browser compliance

0.3 s CLM-2041 fetched from the provider portal over allowlisted 443 · 3 files, 2.4 MB · queued as new

02 typed tools · auth

MCP Gateway

One locked front door — every request is checked, scoped, and logged here.

Tool registry 9 typed tools Auth + quotas per-tenant token Webhook dispatch idempotent

0.5 s token kestrel-prod · 4 of 9 tools allow-listed for this state · span 1 logged

03 state machine

Agent orchestration

Chases the provider portals, reads the evidence, and applies the 214 rules.

Orchestrator supervisor Portal intake agent Damage + doc read agent Rule + policy check agent

1.9 s new → ingested → extracted → validating → decided · 11 spans, 0 retries

04 14 modules

Model layer

The reading engines — the same 14 modules you saw in AI Studio.

Capture 5 modules Structure 6 modules Authenticity 3 modules Embeddings search + fraud Model hosting in-VPC GPU, your account

2.8 s 14 modules ran · 6 fields returned · lowest confidence 0.94

05 partitioned

Data layer

Where documents, rules, and the audit trail live — split per customer.

Document store object storage Vector index search + similarity Rule sets Excel · versioned Audit log append-only

3.2 s document + crops stored · rule set v2.4 read · audit row appended

Decision and citations posted back into Kestrel Insurance’s claims system waiting

Step 01 — Where each customer meets the platform

Tenant surfaces

Every tenant keeps their own intake surface, staff console, and review queue. Nothing about the pipeline leaks between them.

Intake surface for whoever submits the document
Staff console with the review cockpit
Read-only audit browser for compliance

This panel follows the trace. Click any layer to read it on its own.

What the agents replace

Manual today ~11 min each

An adjuster signs into each provider portal, downloads the estimate and photos, then checks 214 rules by hand.

With AuroraAgent 68% straight-through

Agents fetch, read, screen for tampering, and apply the rules. Adjusters see only what got flagged.

Why Private VPC here

General insurance

Policyholder records sit under FSA supervision. Running inside Kestrel’s own account means no third-party data-processing agreement was needed.

Rule set214 rules · v2.4

Document partition1.2M documents

Throughput~42k claims / month

Data residencyap-northeast-1

Rules, documents, embeddings, and audit trails are partitioned per tenant. A tenant token cannot reach another tenant’s data.

Deployment

Documents at restKestrel’s S3 bucket, Kestrel’s KMS key

Model callsStay inside the VPC — nothing leaves the account

EgressOutbound 443 to provider portals only, allowlisted

UpdatesKestrel approves each release before it deploys