Ask OCEAN…
SJ

How OCEAN works

Live agent

The whole system on one page: what it is, the underwriter flow, what runs on AWS, the models, and the data and MCP connectors. Deployed in Alvio's AWS (us-west-2).

1 · What this is

An MVP demo of an agentic underwriting platform, built for the OAK Global / OCEAN interview. It shows how risk and capital decisions could be orchestrated on AWS Bedrock AgentCore. A submission comes in (a broker slip, a form, or a threat event), an event fires, and a durable workflow convenes a committee of specialist AI agents. They pull live external risk signal, run deterministic pricing, accumulation and capital engines, and produce a fully-cited recommendation a human approves.

Three principles run through it. Numbers come from deterministic engines (never the model, so they can't be hallucinated). It is recommend-only (no bind tool exists, a Guardrail blocks any execution claim, and a human decides). Every step is traced and every source is labelled mock vs real-time.

Note: a demonstration for an interview, not affiliated with or built by OAK Global.

Real

The full AWS stack, the agentic orchestration, the deterministic engines, and the live external feeds (GLEIF, USGS, CISA KEV, EPSS, web).

Simulated

Only the book itself (policies, claims, appetite), a representative synthetic dataset in DynamoDB. Swap in a carrier's real book in the same schema and everything above runs unchanged.

2 · The underwriter flow

The evaluation sequence: eligibility first, then the three parallel assessments, then a human. Click any step to see what it does. Try step 1 live on any real company.

Click any step to see how it works and the real systems behind it.

1
pullsslipexposure fileloss run

slip is the deal-terms document.

exposure file is the schedule of what is insured, and where.

loss run is their history of past claims.

A broker sends the deal as documents, not a tidy table. We read them and turn them into structured data automatically (the Ingest tab).

2
pullsGLEIFOFAC

GLEIF is global registry of legal entities; confirms the company is real.

OFAC is the US sanctions list.

The cedent is the insurer asking us to cover part of their risk. We confirm they are a real, registered company and not sanctioned. A hard gate: it can stop here.

3
pullscat modelexposure profile

cat model is software that simulates catastrophes to estimate losses.

exposure profile is what they insure, broken down by size.

We estimate how bad losses could get. Exposure rating works from what they insure, so we can price even with no past-claims history. PML is the worst-case loss at, say, a 1-in-100-year level.

4
pullsour bookappetiteLloyd's RDS

our book is the deals we already hold.

appetite is the limit we set ourselves per area.

Lloyd's RDS is Lloyd's standard disaster scenarios.

Accumulation is how much total risk piles up in one place; one storm or one cloud outage can hit many deals at once.

5
pullsrate benchmarks

rate benchmarks is what the wider market charges for similar cover.

We work out a fair premium. Rate on line is the premium as a percentage of the cover. Terms include surcharges, sub-limits, and reinstatements (buying the cover back after a loss).

6
pullscapital model

capital model is our Solvency II capital calculation.

Regulators require us to hold capital. SCR is the capital we must hold; the solvency ratio is what we have versus what we need.

7
pullsthe decision note

the decision note is the written underwriting produced as the output.

The recommendation (accept, decline, refer, or accept-with-conditions) plus a written underwriting note. A licensed underwriter approves and binds. To bind means to legally commit the cover.

No book needed to start. Steps 1 to 3 and 5 run on the cedent's own submission plus public data (GLEIF, OFAC, hazard feeds, market benchmarks). We exposure-rate the risk, which does not need our own loss history, so a brand-new reinsurer can underwrite from day one. Our book in step 4 simply grows with every deal.

Try step 2 on any real organisation

live, public GLEIF registry, no proprietary data

3 · What runs on AWS (the architecture)

The full event-driven flow, then every AWS component it uses. Each card opens the live console, click to prove it is real.

Provenance AWS (deployed) Engine Real-time Sample book
Intakeslip · form · CSV · threat event
EventBridgeTriggers · event bus
Step FunctionsOrchestration · ocean-assess
AgentCore Runtimethe committee
Guardrailrecommend-only gate
DynamoDBdecision + audit trail
⌄

Inside AgentCore Runtime, the underwriting committee

Leadeligibility → synthesis
3 specialistspricing · exposure · capital (parallel)
Bedrock · LLMreasoning only
Memory (STM)multi-turn
⌄
Tools, AgentCore Gateway (MCP · Cognito-JWT) · 17 tools

Engines - deterministic math over the book

price_treatyquery_accumulationcyber_accumulationcapital_impactcapital_positionscenario_exposureparametric_triggerexposure_summaryrds_scenariossearch_guidelines (RAG)

read the sample book in DynamoDB (treaties · claims · appetite · scenarios)

Live feeds - fetched real-time, keyless

gleif_entity → GLEIFsanctions_screen → OpenSanctionscyber_threat_signal → CISA KEV + EPSSusgs_seismic → USGSweb_search → Web

never the model's memory, pulled at request time

AgentCore Observability, OpenTelemetry traces across every step (CloudWatch GenAI)
Recommend-only: there is no bind/execute tool, and the Guardrail blocks any output claiming to bind, commit capacity or move money, a human approves every decision.

Amazon EventBridge

Event backbone, RiskSubmissionReceived and CyberThreatObserved events trigger assessment.

AWS Step Functions

Durable, auditable orchestration (ocean-assess): Normalize → AssessAndPersist → EmitCompleted, with retries.

Bedrock AgentCore Runtime

Hosts the agents, ocean_uw (underwriting) and ocean_copilot (analytics assistant). Plans, calls tools, reasons.

AgentCore Gateway (MCP)

Exposes the analytics engines + live feeds as 17 MCP tools the agents call over JWT-authed MCP.

AgentCore Memory (STM)

Per-session multi-turn memory for the copilot, resolves follow-ups against earlier turns.

Bedrock Guardrails

Recommend-only policy enforced as an auditable post-hoc check on every output.

AgentCore Observability

OpenTelemetry traces of every reasoning step + tool call (CloudWatch GenAI dashboard).

Amazon Bedrock

Managed foundation model, the reasoning behind the agents (extended thinking + structured output).

Amazon DynamoDB

The book (ocean-book: policies, claims, appetite, scenarios) + decisions (ocean-decisions, audit trail).

AWS App Runner + ECR

Hosts this Next.js app; ECR stores the container images.

4 · The models

Every figure in a recommendation comes from these engines, the LLM owns only the verdict and rationale. Click a model for its type, inputs, outputs, the research it is built on, and to run it live.

The quantitative models that run during a submission. Mostly deterministic actuarial / regulatory models, not ML (the ML is the LLM and EPSS upstream). Click any model for its type, inputs, outputs, the research it is built on, and to run it live as a real MCP call on your own inputs.

5 · Data & MCP connectors

Where the data comes from, how it is wired in as MCP servers, and the full catalog of tools the agents call, each runnable live.

Live public APIs (real-time, keyless)

Sample book (synthetic, in DynamoDB)

  • · Treaties, claims, appetite, scenarios (table ocean-book)
  • · Underwriting guidelines (RAG corpus, Titan embeddings)
  • · The deterministic engines compute over this book

Production would add licensed data via adapters: cat models (e.g. Moody's RMS, Verisk), parametric indices (e.g. Descartes), exposure / geocoding.

MCP servers (the connectors)

Each data source an underwriter uses is a connector (an MCP server). Day-one means public or our own engine, usable with no proprietary data; licensed means a plug-in enabled on engagement. Live servers are wired today (most are runnable in Data → MCP tools); stub servers are on the map, not built yet.

external = outsourced real-world API our engine = deterministic code book = the carrier's book (DynamoDB)

Eligibility & counterparty

GLEIF LEIentity verification
externalday-onelive
OFAC sanctions (web screen)sanctions / adverse-media
externalday-onelive
Ratings AM Best / S&Pcedent financial strength
externallicensedstub

Document intake

Submission ingestion to OEDslip / spreadsheet to structured
our engineday-onelive

Exposure & cat modelling

Cat model RMS / Verisk / Cotalitymodelled loss on exposure
externallicensedstub
Exposure-rating curves (our engine)ILF / MBBEFD pricing
our engineday-onelive

Hazard

USGSearthquake / seismic
externalday-onelive
NOAA stormwindstorm / surge
externalday-onestub
CISA KEV + EPSSlive cyber threat
externalday-onelive

Accumulation & capital

Accumulation + Lloyd's RDSportfolio aggregation
bookday-onelive
Solvency II SCRcapital position
bookday-onelive

Market & benchmark

Rate-on-line indexpricing benchmark
externalday-onestub
PERILS industry lossindustry benchmarks
externallicensedstub

Broker placement

Lloyd's PPL / Whitespacesubmission feed
externallicensedstub

MCP tool catalog

17 runnable · AgentCore Gatewayclick Run to call the real MCP tool, same path the agents use

Eligibility & KYC

gleif_entityVerify a cedent in the live GLEIF LEI registryGLEIF · live
sanctions_screenSanctions / adverse-media screenOpenSanctions / web · liveruntime-local
get_cedent_historyPrior decisions + relationship for a cedentMemory → DynamoDBruntime-local
claims_summaryHistorical claims/losses, total incurred + largestMCP → DynamoDB book

Hazard (live signal)

usgs_seismicEarthquake history near a place (M4.5+)USGS · liveruntime-local
web_searchWindstorm/flood outlook, market conditions, newsWeb · liveruntime-local
cyber_threat_signalInsured's stack vs actively-exploited CVEs + exploit probabilityMCP → CISA KEV + EPSS · live

Pricing, capital & parametric

price_treatyTechnical price for a cat-XL layer (exposure + experience rating)MCP → pricing engine
query_accumulationMarginal 1-in-100 PML vs zone appetiteMCP → accumulation engine
capital_impactMarginal diversified Solvency II SCR + solvency ratioMCP → capital model
parametric_triggerIndex-triggered payout (Horizon / Descartes-style), pays on the index, no loss adjustmentMCP → parametric engine
search_guidelinesSemantic search over underwriting guidelines (cited)MCP → Titan RAG

Technology ecosystem & portfolio analytics

cyber_accumulationShared-dependency accumulation (cloud/software for cyber; region/grid/operator for data centres) vs appetiteMCP → dependency engine
scenario_exposureSystemic scenario (Cloud Down / ransomware / grid outage / data-centre region failure) aggregateMCP → scenario engine
cyber_concentrationALL shared-dependency concentrations in one callMCP → concentration engine
exposure_summaryConcentration across all zones vs appetiteMCP → portfolio engine
capital_positionCurrent own funds, SCR, solvency ratio, headroomMCP → capital model

Data lookups

rds_scenariosRealistic Disaster Scenarios (Lloyd's RDS) net loss vs appetiteMCP → RDS engine
appetite_limitsAll appetite limits + utilisation (zones + cyber dependencies)MCP → DynamoDB book
treaty_lookupFull detail of one in-force treatyMCP → DynamoDB book
cedent_portfolioAll treaties + claims for one cedentMCP → DynamoDB book