An autonomous research agent that scans literature, tests hypotheses on population data, and drafts journal-aware manuscripts — with human review at every stage. Methods and preprints are published in the open.
Each stage produces artifacts the next stage can audit. The pipeline accelerates literature review and analysis — it does not replace scientific judgment.
Scans PubMed and related sources for gaps, contradictions, and under-tested associations. Outputs structured hypotheses with citations — ready for statistical testing, not for automatic publication.
Runs reproducible statistics on NHANES and BRFSS cohorts: logistic regression, trend analysis, chi-square tests, and ROC curves where appropriate. Results are logged for human review before any claim enters a draft.
Produces a journal-aware manuscript draft with real citations pulled from the literature stage. Tone, section structure, and disclosure language follow the target venue — then a human editor takes over.
Human-in-the-loop. Zenlo Research proposes analyses and drafts; scientists and clinicians decide what is sound, what to run next, and what gets submitted. We do not present agent output as peer-reviewed fact.
15 clinical patterns. 8 benchmarked across 5 AI models on 4,018 NHANES adults (F1 up to 0.963). Per-pattern NHANES audits published for 2 (Insulin Resistance, Metabolic Syndrome); the rest are structural CHAI cards, validation in progress.
Every major claim links to a public source. No gated PDFs — open science for open decisions.
Five frontier models benchmarked on a real NHANES cohort — methodology, F1 comparison, and limitations published for independent review.
Submitted to JAMIA Open — we will add a public link when a DOI is available.
Plain-language walkthrough of the benchmark design and what we learned running five vendors on the same patient cohort.
Start with the medRxiv preprint for the full methods, or the Medium article for a shorter overview. Clinical pattern audits for Labs live at /labs/audits.