Grae SI

Sandbox profile

Safety narrative drafts

High intensity

Safety writers draft case narratives from source documents. A reviewer still owns the case. The model runs in the safety VPC.

Vertical
Pharma
Scale
Enterprise
Organization
Kepler Biopharma
Shape
Global pharma safety operations, case processing in two regions.

The business problem

Case volume outruns the writers, and the narrative has to follow the source documents: reporter words, labs, and the company assessment. Sending a case to a public API is not compatible with pharmacovigilance practice or with privacy commitments to affiliates. The team wants a first narrative inside the safety database, with the medical reviewer accountable for what is submitted.

What Grae SI recommends

Family

Qwen 2.5 7B Instruct as the default. BioMistral 7B or Meditron only if counsel accepts the license.

Size

7B LoRA. Not a 70B cluster.

Why this size

Narratives are structured and the source packet is the authority. A domain biomedical model can help terminology, and it can also block you if the license forbids the deployment you need. Start with a 7B general instruct model, fine-tuned on approved historical narratives, and treat a domain base as a second experiment with the license printed on the card.

License

Restricted biomedical licenses are flagged in the catalog. Unknown licenses block redistribution until someone records the terms.

Data controls

PHI and reporter data

Cases contain patient and reporter identifiers. Training uses a de-identified, closed-case extract. Live inference is inside the safety zone and is still personal data.

De-identification

Patient, reporter, and site identifiers removed or tokenized for training. Relative days may replace calendar dates when the medical content allows it.

Access

Case processors and medical reviewers on that product. Region-based access follows the safety system’s existing roles.

Residency

EU cases stay in the EU safety VPC. US cases stay in the US VPC. No combined training set until privacy and the DPAs allow it.

Retention

The submitted case follows GVP retention. The adapter is tied to a case-definition version and retired with a change-control record.

Risk classification

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

  • The narrative can influence a safety submission and describes patient harm.
  • A person reviews before submission. That control is required, and it does not drop the tier to medium.
  • Automated submission would be out of policy for this sandbox.

Compliance requirements to prepare for

ICH E2A and GVP

Serious-case expectations and the company’s pharmacovigilance system. The model may draft. A qualified person remains accountable for the submitted narrative.

ICH E6(R3)

In scope when the case comes from a trial and the narrative supports safety reporting under GCP. Sponsor oversight does not move to the model.

21 CFR Part 11 and GxP

Electronic records and signatures in the safety database. Change control for the model version sits in QA.

HIPAA, GDPR, DPDP

Whichever applies to the data subject. A global company should assume more than one, and keep residency split until a transfer basis exists.

ISO/IEC 42001 and 27001

AI management and security controls for the safety zone. SOC 2 if a partner audit asks, as evidence rather than a Grae SI attestation.

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.

Source fidelity

At least 95% of clinical facts in the draft appear in the source packet, on 120 cases

The narrative is not a place to reconcile conflicting documents. Conflicts should be flagged, not smoothed over.

Seriousness and outcome

No change to seriousness or outcome versus the source assessment

Those fields are not the model’s to invent.

Identifier policy

Training extract passes a 50-row identifier review with zero findings

Live cases stay identified inside the safety tool. Training data should not.

Human sign-off

100% of submitted narratives have the reviewer’s recorded approval

High-tier gate. The eval report lists it as a release condition.

Deployment architecture

DeploymentAir-gapped-capable safety VPC

The safety database never calls the internet for a narrative. Training and serving run in the safety VPC. A one-way drop can move a released bundle into a stricter zone if QA requires a disconnected host.

  1. 01

    Safety database

    Source documents and the case workflow. Roles already exist.

  2. 02

    De-identification job

    Produces the training extract. Does not write back into the case.

  3. 03

    Grae SI

    Job spec, eval, model card, and the offline bundle.

  4. 04

    Serving

    vLLM on a private subnet. Optional second copy via the offline bundle for a disconnected review environment.

  5. 05

    Reviewer

    Edits and signs in the safety database. Submission is unchanged.

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

    Two regional VPCs. A disconnected bundle is available if QA wants a host with no route.

    Buyer effort after the runtime is up: 3–6 person-weeks

  2. Layer 2

    Data ingestion and prep

    Shared

    Grae SI shows what the safety or labeling corpus needs. Your RIM or safety database produces the extract.

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

  3. Layer 3

    Data controls and governance

    Shared

    Subject and reporter identifiers are masked before training. Part 11 retention of the signed record is yours.

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

  4. Layer 4

    Model selection and training

    Grae SI provides

    Default Qwen 2.5 7B. Domain bases appear in the catalog with license restrictions visible before the job starts.

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

  5. Layer 5

    Evaluation and validation

    Shared

    High-tier suite, source-fidelity checks, and a human sign-off record.

    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

    Drift is a safety-operations question: new products, new forms. The registry records the version. The signal detection process stays in pharmacovigilance.

    Buyer effort after the runtime is up: 2–4 person-weeks

  9. Layer 9

    Application and integration

    You own

    The draft returns to the safety system, the med info tool, or the eTMF. Those systems remain yours.

    Buyer effort after the runtime is up: 4–10 person-weeks

  10. Layer 10

    Organizational validation and sign-off

    You own

    QA, drug safety, and privacy sign the intended use and the change-control record.

    Buyer effort after the runtime is up: 4–8 person-weeks

Buyer effort for this profile

16–26 person-weeks for the first product family

The second product family should be closer to 6–10 person-weeks if the VPC and the SOP already exist. The first one pays for the quality path.

  • Drug safety
  • QA validation
  • Privacy
  • Safety-system engineer
  • Regional data protection

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

10–14 weeks for the first product

Includes QA change control. A custom build with a new vendor assessment is often longer than two quarters.

Cost versus a cloud API

Not a realistic option

Case narratives with identifiers are out of policy on a multi-tenant API. The comparison that matters is fully loaded contractor cost: illustrative $80–$150 a narrative in writer time versus a draft that still needs review.

Writer time

20–40% less time on the first draft, if the fidelity gate is met

Illustrative, assuming reviewers still read the source. Not a measured productivity study.

Compute

$2k–$5k a month across two regions

One GPU node per regional VPC, reserved. Training is burst, serving is steady.

Validation writing

About 4–6 person-weeks saved on the first packet

The technical packet is generated. The SOP and the GVP impact assessment are not.

Assumptions

  • Illustrative estimate for this sandbox. It is not a measured customer result.
  • One product family, two regions, historical closed cases available under the privacy basis the company already uses for safety.
  • The model does not determine causality and does not submit to a regulator.
  • A domain biomedical model is used only after the license is accepted in writing.

All sandbox profiles · The full stack