Assembled from the record,
grounded in the record.
Any serious AI system in a regulated setting has three layers. The language model is rented, and everyone rents the same ones. The records are the customer's and always were. What decides whether the output can be trusted is the layer in between, and that is the layer we build.
One thread per patient, one citation per value.
De-identification happens before any model sees a record. Everything after step 2 runs on tokenized data.
What "grounded" means here
An extracted variable is stored with the passage it was derived from and the version of the extraction logic that produced it. A reviewer can open any value and read the sentence behind it. Without that, a real-world evidence dataset is an assertion; with it, it is reviewable.
What "escalation" means here
The system is explicit about what it may conclude alone and what it must hand to a person: low-confidence extraction, contradictory documentation, and anything with a safety dimension. The escalation rules are configuration, not model behavior, so they can be reviewed and audited.
Six properties we hold ourselves to.
1 · De-identification before inference
PHI is stripped and tokenized in our layer, not at the model. The token preserves a patient's continuity across encounters so longitudinal analysis is possible, while the identity stays behind. The model receives the clinical content without the person.
2 · Provenance on every value
A finding that cannot be traced to a source passage is not a finding. Every extracted variable carries its citation and the version of the logic that produced it.
3 · Versioning
Extraction logic and cohort definitions are versioned. An analysis records which version produced it, so a result can be reproduced and a change in method is visible rather than silent.
4 · Explicit escalation
What the system may conclude alone and what it must route to a human is configuration, not model temperament. Low confidence, contradictory source material, and anything with a safety dimension escalate by rule, and the rule is inspectable.
5 · Access logging and authentication
Multi-factor authentication, and access to records logged at the level a HIPAA-aligned control environment requires. Built in from the first commit rather than added when someone asked.
6 · Aggregate-only output to manufacturers
Population-level, de-identified, under the customer's terms. Patient-level records never leave the customer's governance, and nothing returned is designed to be re-identifiable.
Designed for AWS, and for the customer's own account.
Model access through Amazon Bedrock, so inference runs inside the same account boundary and under the same enterprise data terms as the rest of the environment.
- Managed model access. Bedrock rather than a consumer API, for the data-handling terms and for the ability to change models without changing the application.
- Export-based ingest. No EHR API dependency. A scheduled report or file drop is enough.
- Deployment in the customer's account where their governance requires that the records not leave it.
- Model-agnostic. Nothing above the model depends on which model is behind it, including a self-hosted open-weights model where a customer requires one.
A governed, auditable thread through the clinical record, from encounter to evidence.
The model is the least important part of it. The citation, the token and the escalation rule are the parts that make the output usable by someone accountable for it.