Grae SI

Sandbox profile

Translational portfolio briefs

Medium intensity

Portfolio leads ask questions across study reports that cannot be pasted into a public model. Answers cite the report and a scientist checks them.

Vertical
Life sciences and biotech
Scale
Enterprise
Organization
Arcadia Biologics
Shape
Commercial-stage biologics company, a translational group of about 80.

The business problem

Study reports, investigator brochures, and internal memos sit in a document system. New leads re-read the same thousand pages. A hosted assistant would require uploading unpublished human data. The group wants a private reader that answers with a citation, and a person who can say the answer is wrong.

What Grae SI recommends

Family

Qwen 2.5 Instruct

Size

7B, with retrieval over the approved document set rather than a larger model that memorizes it.

Why this size

The facts live in the documents. Retrieval plus a 7B model is cheaper and easier to validate than a 70B model that still needs citations. The adapter, if any, is for tone and refusal, not for hiding the corpus inside weights.

License

Same catalog rules. Retrieval does not relax the license of the base model.

Data controls

Human data

Reports may contain subject-level tables. Those tables are indexed only if they are already coded to the company’s standard, and subject identifiers are not embedded in the prompt log in the clear.

De-identification

The index uses coded ids. A preprocessing step drops names and site-level rare events the privacy office flags.

Access

Portfolio role in the document system. The model service trusts that role and does not invent a second ACL that can drift.

Residency

Index and model in the R&D VPC in the region where the studies are held.

Retention

Answers that are copied into a regulated document follow that document’s schedule. The chat log for exploration expires in 30 days unless a user files it.

Risk classification

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

  • The user is a scientist, the output is an answer with a citation, and it does not change a protocol or a batch.
  • Subject-level human data in the corpus keeps the tier off low.
  • Using an answer in a regulatory submission without a documented review would move the use to high.

Compliance requirements to prepare for

ICH E6(R3)

When the source is a clinical study report. The model is not a change to the protocol or to trial conduct.

GDPR, HIPAA, or DPDP

Depends on whether the indexed tables are personal data in that jurisdiction. Coded research data can still be personal. Privacy makes that call.

ISO/IEC 27001 and 42001

Security of the index, and an intended-use record that says answers are advisory.

GxP

Out of the quality system until someone pastes an answer into a regulated record. The SOP should say what happens at that moment.

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.

Citation accuracy

At least 90% of factual sentences match the cited passage on a 100-question set

An answer without a true citation is a miss.

Refusal

100% refusal on a 25-question set that asks for a protocol change or a batch disposition

Scope control.

Access mismatch

A user without document permission gets no passage from it, tested on 20 probes

The model must not become a side door around the document ACL.

Staleness

Index rebuild within 7 days of a new approved report

A retired memo should not answer a portfolio question.

Deployment architecture

DeploymentPrivate R&D VPC

The document system stays the system of record. A private index and a 7B runtime sit beside it. No document bytes go to a model API outside the account.

  1. 01

    Document system

    Permissions remain here.

  2. 02

    Private index

    Built inside the VPC from approved reports. Grae SI does not host the index.

  3. 03

    7B runtime

    vLLM, private endpoint, answers must quote a passage id.

  4. 04

    Reader

    A simple internal page. Filing an answer into a regulated doc is a separate, human step.

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

    A private VPC or an air-gapped quality zone. No training call to a public API.

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

  2. Layer 2

    Data ingestion and prep

    Shared

    Indexing approved reports is the company’s pipeline. Grae SI’s prep step covers a sample and the data-class checklist.

    Buyer effort after the runtime is up: 3–6 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

    Shared

    Grae SI selects and packages the 7B model. The retrieval index is the company’s, because it is a copy of their reports.

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

  5. Layer 5

    Evaluation and validation

    Shared

    A tier-matched suite, thresholds, and an exportable report. The labeled set, the acceptance call, and any statistics your quality system requires on top.

    Buyer effort after the runtime is up: 1–3 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

    Registry lineage: version, dataset, eval result, and the package that was released. Live traffic checks, drift alarms, and the decision to pull a version.

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

  9. Layer 9

    Application and integration

    You own

    An internal page that respects document permissions. Not a new portfolio platform.

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

  10. Layer 10

    Organizational validation and sign-off

    You own

    QA releases the model version inside your GxP process. The Grae SI report is an input to that record, not the release.

    Buyer effort after the runtime is up: 3–8 person-weeks

Buyer effort for this profile

8–14 person-weeks

Retrieval quality dominates. The language model is the smaller risk and the smaller cost.

  • Translational lead
  • Knowledge-engineer or the person who already runs search
  • Privacy
  • One platform engineer

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

6–9 weeks

Assumes documents are already in one system with permissions. If they are on laptops, add a migration that this estimate does not include.

Cost versus a cloud API

About $8k–$20k a month avoided

Illustrative: 40 scientists, heavy daily use, long reports. Uploading unpublished study reports to a shared API is also out of policy, so the avoided cost is partly hypothetical.

Compute

$1.2k–$2.5k a month

One GPU node for the 7B model plus a modest CPU index.

Search time

30–50% less time to a cited answer on the gold questions

Illustrative against unaided re-reading. Not a timed user study from a customer.

Accuracy target

90% citation match

Measured on the 100-question set before go-live.

Assumptions

  • Illustrative estimate for this sandbox. It is not a measured customer result.
  • The document ACL is the access control. Building a second permission model is out of scope and out of the estimate.
  • Answers are not pasted into a submission without a documented scientific review.
  • This profile is a reader. Digital twins of patients, disease, or bioreactors are a different profile.

All sandbox profiles · The full stack