Sandbox profile
Digital twins for pharma and life sciences
High intensityBuild patient, disease, and manufacturing twins in-house. Grae SI governs the model, the evidence, and the offline package. It is not a twin-as-a-service, and it does not compete with a vendor that sells control-arm outcomes as a service.
- Vertical
- Pharma and life sciences
- Scale
- Enterprise
- Organization
- In-house twin program
- Shape
- A pharma, biotech, or CRO team building twins it owns. A 40-person biotech starts with one twin. A large sponsor may run all three.
The business problem
Teams want a quantitative stand-in for a patient journey, a disease trajectory, or a bioreactor, trained on data they cannot ship to a service bureau. The useful questions are specific: can a twin support a smaller control arm, can it forecast response in a defined population, can it warn that a batch is drifting before release testing. Those models are not chatbots. They are small scientific models with a credibility story a regulator or a QA unit can read. Renting the twin means the data, the assumptions, and the lock-in sit with the vendor. The team wants the weights, the dataset lineage, and the right to run with the network unplugged.
What Grae SI recommends
Family
A compact domain model you bring, plus a 7B instruct model only for the narrative around it
Size
Twin: low millions of parameters, or a gradient-boosted / state-space model small enough to sit on a CPU node. Narrative: Qwen 2.5 7B, optional, and not the twin.
Why this size
A 7B language model does not simulate a bioreactor or a longitudinal patient. The twin is a model the scientific team already knows how to train, imported through bring-your-own-model, then packaged, evaluated, and registered in Grae SI. The language model, if used at all, drafts the credibility narrative from statistics the twin has already locked. It does not invent endpoints. A small biotech can stop after one twin on a single CPU or GPU node. An enterprise sponsor repeats the same pattern per context of use, rather than training one universal model.
License
The twin’s training code and any pretrained encoder need a license the company can use in a regulated submission. The 7B narrative model’s license is separate and is printed on its own card. Neither license is a regulatory clearance.
Data controls
Longitudinal patient data
Patient twins need repeated measures: visits, labs, outcomes, and the covariates the context of use names. That is PHI or coded clinical-trial data. It stays in the study zone.
De-identification
Training extracts are coded. Dates can be study-relative. The credibility plan states whether the data are identifiable, coded, or de-identified, because that choice changes HIPAA, GDPR, and the trial agreement.
Manufacturing data
Bioreactor twins use sensor series, setpoints, and release results. Those are trade secrets more often than personal data. They still do not go to a multi-tenant training API.
Access
The scientific owner and QA. A CRO sees a sponsor’s twin data only under the study contract. Patient-level rows are not a demo feature in a shared sandbox.
Residency and retention
The zone follows the trial or the plant. Retention follows the TMF or the batch record, and the model version is cited there. When a study contract requires destruction, the registry entry says which artifact was destroyed.
Risk classification
Intensity is a planning tier for evidence. It is not a certification and not a regulatory classification.
- A patient twin used to justify a smaller control arm can change a trial and, later, an evidence package.
- A manufacturing twin used before batch disposition can affect product release. Even a “warning only” use is high until QA says the disposition process cannot see it.
- A disease model used only inside a research notebook, with no protocol impact, can be argued down to medium. This flagship profile assumes the team wants a context of use that could support a decision, and it stays at high until the context of use is narrowed in writing.
Compliance requirements to prepare for
FDA AI credibility framework
For drug and biological products, FDA’s risk-based credibility approach asks for a defined context of use, a model-risk read (how much influence, how serious the decision), and a credibility assessment plan matched to that risk. Grae SI stores those three as the validation spine. It does not agree them with FDA.
ICH E6(R3)
When the twin touches a clinical trial: protocol, statistical analysis plan, and sponsor oversight still govern. A twin is not a protocol amendment by itself.
21 CFR Part 11 and GxP
If outputs enter an electronic record that supports a submission or a batch record. Signatures and change control stay in the quality system.
HIPAA, GDPR, DPDP
Patient-level longitudinal data. Manufacturing sensor data may be outside those laws and still confidential. Say which is which in the data-class record.
ISO/IEC 42001 and 27001
The management system around an AI model that now has a context of use, not only a notebook.
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.
Prediction error
Pre-declared margin on a locked test set, for example mean absolute error or survival error within the bound in the statistical analysis plan
The number is study-specific. The requirement is that the bound exists before unblinding or before the model can influence the decision.
Calibration
Calibration slope between 0.8 and 1.2 on the held-out set, or a bound the statistician sets instead
A twin that is sharp and biased is not credible for a control-arm discussion.
Control-arm size
Any claimed reduction, for example 15–30% fewer randomized controls, is a design scenario in the SAP, not an output the model prints
The reduction is only real if the design and the regulator accept the context of use. The sandbox does not claim that reduction has been achieved.
Process-twin warning
Alert precision and recall agreed with manufacturing science on historical batches, before any alert is visible to release QA
An alert that QA can see is already in the quality system.
Narrative lock
The 7B narrative cites only figures present in the locked twin output. 100% on a 30-report probe.
The language model is not allowed to improve the science in prose.
Three twins, built in-house
Unlearn and similar companies sell external control arms and related analyses as a service. This profile is the opposite commercial shape: the sponsor, biotech, or CRO builds the twin, keeps the weights, and runs it in its own zone, including offline. Grae SI is the path that makes that model reviewable. It is not the twin.
Patient twins for trial control arms
A model of the control trajectory for a defined population, used only in a context of use the protocol and the analysis plan name. It can support a discussion of how many randomized controls are needed. It does not replace that discussion, and it does not enroll anyone.
Data. Longitudinal control-arm or natural-history data: repeated outcomes, baseline covariates, and the visit structure. Coded subject data, region-locked to the study.
Model. A compact time-to-event or longitudinal model. Bring it in as the artifact. Size is whatever the statistician can validate, typically far smaller than a language model.
Disease and drug-response twins
A model of progression or response for a defined disease and treatment class, used in research or in a named decision such as dose selection. The context of use says which decisions it may inform. Everything else is out of bounds.
Data. Longitudinal clinical and, where justified, biomarker series. Consent and the protocol have to allow the secondary use. If they do not, the twin waits.
Model. The same bring-your-own path. A pretrained encoder is allowed only with a license that survives a submission and a documented fine-tune on data the company may use.
Bioreactor and manufacturing process twins
A model of a unit operation from sensor history, used first as an offline warning. It becomes part of release only after QA changes the procedure. Until that change, release QA does not see the alert.
Data. Historian tags, setpoints, in-process controls, and historical release results. Usually not personal data. Still confidential, and still not a candidate for a public training API.
Model. A small state-space, LSTM, or gradient-boosted model on the tag set. CPU serving is often enough. Package it like any other Grae SI artifact so the version on the plant floor is the version QA read.
Credibility plan
Context of use
Who uses the output, what decision it may influence, and what it must not do. A research notebook and a control-arm design are different contexts and different evidence burdens. Write the sentence before training the final model.
Model risk
Combine how much the model can sway the decision with how bad a wrong decision is. A warning nobody is allowed to act on is lower influence than a number in the sample-size section of a protocol. The tier in Grae SI stays high until that influence is actually limited.
Credibility plan
The evidence you will collect for that risk: data fitness, prediction error, calibration, subgroup behavior, and what would make you throw the model out. The plan is locked before the confirmatory test is opened.
What Grae SI will not do
It will not tell you the context of use is acceptable to FDA, EMA, or a notified body. It will not sign the batch record. It will not claim a control-arm reduction as a result.
Deployment architecture
Data platforms stay where they are. The twin trains there. Grae SI packages the weights, the eval, and a runtime that can be copied to a host with no route. A language model is a sidecar for narrative drafts, on the same zone, and it reads only the locked metrics.
01
Longitudinal store
EDC, labs, or the historian. Stays in the zone. Exports are coded.
02
Twin training
The scientific stack the team already uses. Brought into Grae SI as the model artifact, not rewritten as a chatbot job.
03
Credibility record
Context of use, model risk, plan, dataset lineage, and the metric card. This is the Grae SI validation step.
04
Offline runtime
The twin serves in-zone or from the air-gapped bundle. No call to a twin vendor.
05
Narrative sidecar
Optional 7B model. Input is the locked metric table. Output is a draft a person edits.
06
Sign-off
Statistician, scientific owner, and QA. A regulator, if any, sees the packet the company chooses to file.
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
- You own
Layer 1
Infrastructure and compute
CPU node for many twins. A GPU only if the scientific model needs it. The zone already exists for the trial or the plant.
Buyer effort after the runtime is up: 2–6 person-weeks
- You own
Layer 2
Data ingestion and prep
Longitudinal curation is the scientific team’s. Grae SI checks a sample and records the data classes. It does not build the SDTM pipeline.
Buyer effort after the runtime is up: 4–10 person-weeks
- Shared
Layer 3
Data controls and governance
Coding, residency, and contractual destruction. Grae SI holds the checklist and the version that must be destroyed.
Buyer effort after the runtime is up: 2–4 person-weeks
- Shared
Layer 4
Model selection and training
You train the twin. Grae SI’s bring-your-own-model path records the family, the size, the license, and the job that produced the artifact. The optional 7B narrative model comes from the catalog.
Buyer effort after the runtime is up: 3–8 person-weeks
- Shared
Layer 5
Evaluation and validation
High-tier gate plus the credibility plan’s metrics. Grae SI exports the report. The statistician defines the margins.
Buyer effort after the runtime is up: 4–8 person-weeks
- Shared
Layer 6
Agents and orchestration
Compliance agents classify the use case. They do not orchestrate the twin’s scientific code.
Buyer effort after the runtime is up: 1–2 person-weeks
- Grae SI provides
Layer 7
Deployment and serving
Offline bundle so the twin runs without a vendor account. You move it into the zone.
Buyer effort after the runtime is up: 1–2 person-weeks
- Shared
Layer 8
Monitoring and drift
Drift against new sites, new batches, or a protocol amendment is yours to watch. The registry says which version was current.
Buyer effort after the runtime is up: 3–6 person-weeks
- You own
Layer 9
Application and integration
A notebook is enough for a research context. A control-arm or a batch-alert context needs a real integration, and that integration is yours.
Buyer effort after the runtime is up: 4–12 person-weeks
- You own
Layer 10
Organizational validation and sign-off
QA or the sponsor’s clinical leadership signs. FDA interaction, if you seek one, is yours. Grae SI does not request a meeting and does not represent you.
Buyer effort after the runtime is up: 4–10 person-weeks
Buyer effort for this profile
12–20 person-weeks for a first research twin at a small biotech. 30–50 person-weeks for a first decision-supporting twin at a sponsor or CRO.
A small company should run one context of use. An enterprise program repeats the package per twin. The second twin reuses the zone and the SOP, and the estimate drops by roughly a third if the first sign-off path worked.
- Scientific owner (clinical, quantitative, or manufacturing science)
- Statistician or process statistician
- QA
- Data engineer for the longitudinal extract
- Privacy, if patient-level
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
One research twin in about a quarter. A decision-supporting twin in two quarters.
Assumes the longitudinal extract already exists. If the data are still in spreadsheets, this estimate is wrong and the data work dominates.
Cost versus renting a twin service
Service contracts for external control arms are often in the high six or seven figures per program. In-house compute for a compact twin is illustrative $2k–$8k a month plus the people above.
The service price is a market range, not a quote, and not every program can replace a service with an internal model. The point of the comparison is ownership and data movement, not a promise of savings.
Cost versus a cloud LLM API
The API does not do this job
A chatbot call does not produce a calibrated longitudinal twin. Budget the scientific model, not a token price.
Validation writing
About 4–8 person-weeks saved on the technical credibility packet
Context of use, lineage, metric tables, and the model card are assembled. The statistical analysis plan and any FDA briefing book are not.
Control-arm scenario
A design target of 15–30% fewer randomized controls, only if the context of use is accepted
Illustrative scenario for planning sample size. It is not an observed reduction, and it is not available for every endpoint.
Manufacturing
Earlier warning on historical batches, measured by the precision and recall QA agreed
No dollar savings are claimed. A plant should not book a yield number until the alert has a locked test.
Assumptions
- Illustrative estimate for this sandbox. It is not a measured customer result.
- Grae SI does not sell twin outcomes, control-arm patients, or a bioreactor model as a service. The company trains and owns the twin.
- The language model is optional and cannot change a locked metric.
- FDA’s credibility framework is a way to organize evidence for a drug or biological product model. Using it in the product is not an FDA agreement, a qualification, or a clearance.
- Small-company plan: one twin, research context, low end of the effort range. Enterprise plan: a decision-supporting context, high end, then reuse.
- Longitudinal data are already curated. Collection and SDTM build are outside the estimate.