Grae SI

Sandbox profile

Underwriting file briefs

High intensity

Underwriters get a structured brief of a submission. The model does not price, bind, or decline.

Vertical
Insurance
Scale
Enterprise
Organization
Northwind Assurance
Shape
National P&C carrier, personal and small commercial lines.

The business problem

Submissions arrive as applications, loss runs, and notes. Underwriters spend the first pass assembling a picture. A vendor score that declines a risk is a different product, and a regulated one. Northwind wants a brief a person reads before using the existing rating engine. The brief must not be a shadow price.

What Grae SI recommends

Family

Qwen 2.5 Instruct or Llama 3.1 Instruct

Size

7B–8B, one private GPU pool in the underwriting VPC.

Why this size

Loss runs are long. A 3B model drops years. A 7B–8B model with a fixed brief schema (exposures, losses, missing items, contradictions) is the smallest that holds the file. Pricing stays in the actuarial system, which is already governed.

License

If Llama’s community license conflicts with a use the carrier wants later, the card says so now, while the use is still a brief.

Data controls

Personal and commercial data

Applications include names, addresses, and sometimes driving or loss history that is personal. Small commercial files include business data the carrier treats as confidential.

De-identification

Training is on closed submissions with direct identifiers removed. Protected-class fields are not features and are stripped from the training text.

Access

The assigned underwriter and referral authority. The model service cannot see files outside that authority.

Residency

US underwriting VPC. A state that requires in-state handling is a separate zone, not a prompt routed overseas.

Retention

Saved briefs follow the underwriting file schedule. Training rows are dropped when the program’s forms change materially, then the model is retrained and re-evaluated.

Risk classification

Intensity is a planning tier for evidence. It is not a certification and not a regulatory classification.

  • Underwriting is a regulated decision even when a person still clicks the rating engine.
  • A brief changes what the underwriter sees. That influence, plus consumer impact, is high intensity.
  • The control is a ban on price, tier, and decline language, tested before release, plus the existing underwriting authority.

Compliance requirements to prepare for

NAIC AI model bulletin

Governance, testing, and third-party due diligence. A private model does not skip the bulletin. It changes who operates the control.

State insurance AI rules

Including unfair-discrimination testing where the state requires it for underwriting models. This brief is in scope of that conversation even though it does not emit a price. Examples include Colorado’s framework and New York’s insurance AI expectations. Map the carrier’s actual states.

SOC 2, ISO/IEC 27001, ISO/IEC 42001

Security, AI management, and the evidence pack. Auditors and rating agencies ask. Grae SI does not certify.

GDPR

If a program writes EU risks, personal-data duties apply and the US VPC is the wrong residency. This sandbox assumes a US book until that changes.

Grae SI helps assemble the evidence. It does not certify HIPAA, FDA, ISO, SOC 2, ICH, NAIC, GDPR, DPDP, or any other framework.

Validation KPIs

These are targets to lock before go-live. They are not measured outcomes from a customer deployment.

No price or decline

0 price, tier, or decline recommendations on a 60-prompt probe

The rating engine remains the only price.

Loss-run fidelity

At least 95% of loss figures match the source on 100 files

A wrong loss total is worse than no brief.

Protected-class silence

Brief does not mention race, religion, or proxies the compliance list names, on a probe set

The training extract removed the fields. The probe checks the model did not invent them.

Human authority

100% of binds go through the existing underwriting authority, unchanged

The workflow test, not a model score.

Deployment architecture

DeploymentPrivate underwriting VPC

Rating and policy admin stay on their current hosts. The brief service is a private endpoint. It cannot call the rating API. That missing wire is deliberate.

  1. 01

    Submission file

    Application, loss runs, notes. Authority model already exists.

  2. 02

    Strip and train

    Identifiers and protected-class fields removed from the training extract.

  3. 03

    7B brief service

    vLLM. Schema with no price field. No route to the rating engine.

  4. 04

    Underwriter desktop

    Brief beside the file. Price is still produced only by the actuarial system.

  5. 05

    Monitoring

    Compliance samples briefs. The registry holds the version that was sampled.

Each box is a component in the customer’s network. Arrows are the order of the path, read left to right on a wide screen and top to bottom on a phone.

Where this sits in the stack

The same ten layers appear on every profile. The mark says whether Grae SI provides the layer, shares it, or leaves it with you. Person-weeks are the buyer’s effort after the runtime is deployed, not a project plan for the whole company.

  • Grae SI providesGrae SI provides the layer
  • SharedShared. Grae SI supports part of it.
  • You ownYou own the layer after install

  1. Layer 1

    Infrastructure and compute

    You own

    Private VPC next to the policy admin host. Prompts do not leave the carrier network.

    Buyer effort after the runtime is up: 1–3 person-weeks

  2. Layer 2

    Data ingestion and prep

    Shared

    A checklist of what to collect, a sample upload, and a readiness score. The pipeline from the system of record into an approved extract.

    Buyer effort after the runtime is up: 2–5 person-weeks

  3. Layer 3

    Data controls and governance

    Shared

    Policy numbers and SSNs are redacted in the prep preview. The carrier still owns the access model and the state privacy notices.

    Buyer effort after the runtime is up: 1–3 person-weeks

  4. Layer 4

    Model selection and training

    Grae SI provides

    7B–8B instruct model. Output schema is part of the job spec so a later change is a new version, not a prompt edit in production.

    Buyer effort after the runtime is up: 0.5–1.5 person-weeks

  5. Layer 5

    Evaluation and validation

    Shared

    High-tier suite plus the no-price probe and a loss-run fidelity set. Compliance samples before release.

    Buyer effort after the runtime is up: 3–5 person-weeks

  6. Layer 6

    Agents and orchestration

    Shared

    Compliance agents that classify the use case and name the frameworks to prepare for. Any production agent that calls tools inside your applications.

    Buyer effort after the runtime is up: 1–2 person-weeks

  7. Layer 7

    Deployment and serving

    Grae SI provides

    Safetensors, GGUF, a model card, and an offline bundle for Ollama, vLLM, or llama.cpp. Moving that bundle into the zone and keeping the host patched.

    Buyer effort after the runtime is up: 0.5–1.5 person-weeks

  8. Layer 8

    Monitoring and drift

    Shared

    The carrier’s model-risk or compliance team samples live briefs. Grae SI does not receive the files.

    Buyer effort after the runtime is up: 3–5 person-weeks

  9. Layer 9

    Application and integration

    You own

    Desktop integration without a wire to the rating engine.

    Buyer effort after the runtime is up: 6–10 person-weeks

  10. Layer 10

    Organizational validation and sign-off

    You own

    Underwriting governance and compliance accept the use. State filings, if any, are the carrier’s counsel.

    Buyer effort after the runtime is up: 4–7 person-weeks

Buyer effort for this profile

18–28 person-weeks

The ban on a shadow price is a design constraint, and it is also what keeps the validation scope smaller than a full rating model.

  • Underwriting operations
  • Compliance
  • Model risk or actuarial governance
  • Security
  • Desktop integration

Quantitative benefits

Illustrative estimate for this sandbox. It is not a measured customer result. Assumptions for this profile are listed under the figures.

Time to a reviewable package

12–16 weeks

Includes a compliance review and a state-mapping memo. A new underwriting score would be a longer, different project.

Cost versus a cloud API

About $25k–$60k a month avoided

Illustrative: 50,000 submissions a month, long loss runs, $0.5–$1.20 a file. Many of those files could not go to a multi-tenant API under the carrier’s security standard.

Compute

$3k–$7k a month

A small private GPU pool sized for business-hours traffic, scaled down overnight.

First-pass time

15–25 minutes less per submission

Illustrative if fidelity targets are met. Underwriters still open the loss run.

Validation writing

About 4–6 person-weeks saved

Technical packet only. The unfair-discrimination analysis, where a state requires one, is additional and is not included as saved.

Assumptions

  • Illustrative estimate for this sandbox. It is not a measured customer result.
  • US book only. EU risks are a different residency design.
  • The model has no path to the rating engine.
  • Protected-class testing follows the states counsel names. The probe in this profile is a start, not that legal analysis.

All sandbox profiles · The full stack