Frekil

The case against most 'AI for clinical research' tools is that they are fluent but not accountable: you get a plausible answer with no trace, and no biostatistician will stake a publication on it. Frekil's answer is to keep the human in the loop at every stage and to keep the data away from the model. Concretely, you review the causal DAG, adjust cohort definitions, and tweak the Statistical Analysis Plan before the SAP Executor runs; the SAP Executor produces R, Python, and SAS code in a sandboxed environment and hands back an audit trail from raw data to final survival analysis. That traceability is what the seed's own user quotes keep coming back to — a senior biostatistician citing 'a transparent, auditable trace from the raw data to the final survival analysis,' a professor of epidemiology citing reproducibility and a complete set of TFLs with the underlying scripts, and a clinical researcher citing code they could independently verify. The pipeline itself is broad: the Literature Review step searches PubMed and your proprietary studies and extracts effect sizes and covariates (the site's example pulls HR 0.74 from one trial and flags adjustment for immortal time bias in another); the Cohort Builder defines temporal windows and automatically detects immortal time bias; the Causal DAG Builder maps effects cited from literature and validates confounders; the Data Extractor maps schemas to OMOP CDM and runs routine missingness EDA; the Report Writer emits a publication-style PDF with audit trail. The company is backed by Y Combinator and publishes methodological content — on empirical calibration and negative controls for p-value reliability, on propensity scores for survival outcomes, and on why payer access is a separate problem from regulatory approval — which signals that it wants to win on statistical trust rather than on marketing. Where it does not fit: it needs structured patient-level data. Teams without EHR, claims, or registry access, or with datasets too small to justify enterprise onboarding, will get little from it. It is evidence-generation infrastructure, not point-of-care decision support — there is no real-time clinical alerting. And per the vendor's own framing, the pipeline is aimed at retrospective observability work (RWE), so laboratory or preclinical research without patient-level data is out of scope. Frekil also does not publish specific AI model names, so if you need to know which foundation model powers the agents, this page cannot tell you and neither can the site. The honest read: for pharma, biotech, and CRO evidence teams with governed clinical data and a rigorous review process, Frekil compresses months of biostatistics and data engineering into days; for everyone else, it is not the tool.