The loop, the harness, the lakehouse. The model sits where it can be swapped.
The lakehouse is memory; the harness is execution and control; the human signature is calibration. The hard part is not making the AI clever; it is keeping everything coherent: what the model sees, what it may do, what humans have signed, and where the trial actually stands. Every turn of the loop re-aligns it.
Six components that keep the system coherent.
Execution, calibration and memory maintain the clinical context together. Beneath: sandboxed execution, six-dimension permissions, identity kept apart from data, so the AI never sees who the patient is.
Rules, thresholds and gates define where the model may work.
Fine-tuned, distilled models for specific clinical tasks.
Red-line tests and regression cases every change must pass.
Who changed what, on what basis, and why.
Subject, event and protocol linked with timeline and lineage.
Handoffs, acceptance and decisions share one form.
A number is the same number everywhere.
Data enters the sponsor's own lakehouse. Connectors, OCR and vision-language parsing bring it in; locally licensed controlled terminology normalises it; quality gates release immutable, versioned snapshots with field-level lineage.
- EDC
- Central labs
- Imaging
- Prior archives
- Connectors
- OCR
- VLM parsing
- MedDRA
- CDISC SDTM
- LOINC
- UCUM
Company-reported recognition accuracy from Raw into downstream layers is 99%+.
Every number is an object, and every object has a source.
What lives in the lakehouse is objects, not text chunks: study, subject and visit form the spine; source batches, artifacts and segments are the entry; every inference is a versioned run; raw observations, facts and derived facts are joined by evidence edges; dictionaries and units normalise facts; issues, review tasks and audit logs cut across everything. Below is the public, simplified view.
- STUDY
- SITE
- SUBJECT
- VISIT
Protocol versions and the visit matrix hang here.
- SOURCE_BATCH
- ARTIFACT
- ARTIFACT_SEGMENT
Every artifact has a hash and locator; every segment resolves to page, table, row.
- INFERENCE_RUN
- RAW_OBSERVATION
- CONCEPT_CANDIDATE
Model, version and prompt version live on the run; extraction, parsing and confidence scores are kept apart.
- EVENT
- FACT
- DERIVED_FACT
- FACT_EVIDENCE
Facts are supported by raw observations; derived facts carry their rule and inputs; SDTM / ADaM are projections, not copies.
- CONCEPT_DEF
- CONCEPT_ALIAS
- CONCEPT_EXTERNAL_MAP
- UNIT_DEF
LOINC, MedDRA and UCUM are external maps; new concept candidates enter the dictionary after review. The dictionary learns too.
- DATA_ISSUE
- REVIEW_TASK
- DECISION
- AUDIT_LOG
Any object can raise an issue or a review; sign-offs enter the ledger; every change keeps old value, new value and reason.
Public simplified view; field-level design is not on this site.
Risk oversight, grounded in the same evidence.
Critical-to-quality factors, quality tolerance limits and key risk indicators run on versioned snapshots. Validation findings feed monitoring, risk signals trigger human review; evidence and rationale are retained and feed rule updates and regression checks.
The co-pilot drafts, cites, analyses, proposes. A human always signs.
Every signature is a calibration: corrections become regression cases, and a threshold cannot change until it passes them. Today that means drafts, answers and records aligned to one context; next, decisions such as dose escalation and observation windows arrive with their full evidence chain attached.
A signature counts only when it carries all six.
Today
Harness and lakehouse are private to the sponsor. The foundation model is called externally behind a vendor-switch layer. Queries, handoffs and alerts share one snapshot ID; project assets stay with the sponsor.
Direction
As task models, eval sets and context accumulate, dependence on any single model, closed models especially, keeps shrinking. A fully private deployment is the goal, not a current claim.