The engine
One engine. Measured, explainable, reproducible.
A cohort's lab panels in. A measured biomarker signature out — with the reasoning shown at every step.
Pipeline
A cohort's panels go in. A measured signature comes back. Eight steps, grouped in three stages — deterministic on the hard rules, LLM only at the edges, explainable all the way through.
Input
Ingest
PDF, lab export, photo, or manual entry becomes structured biomarker data (OCR + extraction).
Normalize
Units reconciled; each value mapped to the Biomarker Library — 101 biomarkers, reference ranges by age and sex.
Engine
Detect patterns
The profile is tested against 15 clinical patterns and 8 interaction types.
Quantify
HOMA-IR and Biological Age computed deterministically; per-axis metrics normalized.
Compare (before → after)
Baseline vs follow-up, with directional meaning (lower / higher / U-shape is better) per biomarker — the cohort shift.
Ground
Every reading checked against the Knowledge Store: population data and literature.
Output
Synthesize
Shifts by axis, significance, drivers, and what moved — assembled into the Signature Report.
Audit trail
Every decision traced and reproducible — the evidence behind the claim.
Per-axis Δ, significance, biomarker drivers, and a full audit trail.
Design
Deterministic first, AI at the edges. The hard rules are computed, not guessed. A model handles only the ambiguous parts — and every call runs with zero data retention.
Computed, not guessed
- HOMA-IR, Biological Age, per-axis indices
- Reference ranges by age and sex
- Pattern thresholds & compound N-of-M logic
- Same input → same reading, every time
Only the ambiguous parts
- Messy or non-standard lab names
- Contested or borderline profiles
- Plain-language rationale for each call
- Zero data retention on every call
What's inside
Six modules, all live
The engine is a stack of validated components — each one running in production today.
Biomarker Library
101 biomarkers, reference ranges by age and sex.
Clinical Pattern Engine
15 validated patterns + 8 interaction types.
HOMA-IR & Bio-Age
Deterministic indices — computed, not guessed.
Shift engine
Before → after with directional meaning per biomarker.
Knowledge Store
Population data + literature for grounding.
Multi-model validation
The same dataset run through five models.
Validated, published, reproducible. Clinical patterns validated on 4,018 NHANES patients and benchmarked across five models; the methodology is published as a preprint. A separate research pipeline runs on public datasets only — isolated from any client data.
See it on your cohort
Send one trial's before/after panels and we return a Signature Report: a measurable profile shift with an evidence trail.