From clinical data to research evidence.
Bring a question and your data. Confirm what it means, build a cohort, run guided statistics and models, and leave with findings you can trace back to their source, without writing code.
Free during early access. No payment, no plan to choose.
Raw records
420 patients, nothing confirmed
Confirmed
Fields carry meaning
Cohort
168 meet the criteria
Result
Readmission by follow-up
One workspace, from question to finding.
Each step confirms the last. The product refuses to run on a guess, and tells you why when it stops.
01
Frame the question
Write it the way you would say it to a colleague. The assistant proposes a structure (population, exposure, comparator, outcome, timeframe) and your original wording is kept exactly as you typed it.
02
Confirm what the data means
Upload CSV or Excel. The profiler infers table grain, field meaning, units, identifiers, candidate outcomes and likely PHI, then asks you to confirm. Nothing downstream runs on a guess.
03
Define the population
Cohort rules stay visible and editable before they run. The count appears only after execution, never as a speculative live number.
04
Run the analysis
A method is recommended with its rationale and the assumptions it will check. Every figure is computed by scipy, statsmodels and scikit-learn.
05
Read the findings
Evidence is synthesised against your original question, with limitations and provenance on every card, and exported as a report that carries them.
Built for work that has to hold up.
Profiling and watchers
Distributions, missingness and data-quality checks, plus deterministic watchers that surface confounding and bias with the numbers behind them.
Reusable cohorts
Rule trees over the confirmed schema. Every run is kept, so you can see how a population changed.
Guided statistics
Assumption checks before the test runs, effect sizes with confidence intervals, and the method recorded alongside the result.
Models with guardrails
Leakage detection, an explicit validation strategy, calibration and feature attribution.
Synthetic populations
Generate statistically plausible patients to rehearse an analysis before touching real records. The synthetic label follows every artifact derived from them.
Traceable export
A report whose every claim points back to the artifact, method and data version that produced it.
What the product will not do.
These are enforced in code, not stated in a policy page.
AI never produces a number
Language models structure questions, infer what columns mean, recommend methods and interpret results. Every p-value, coefficient, metric and cohort count comes from deterministic computation. An interpretation that introduces a figure the analysis did not produce is rejected before you see it.
Association is not causation
Causal wording is flagged, explained and offered an associational rewrite rather than silently accepted. Only a randomised design supports a causal claim on its own, and the workspace says so where it matters.
Protected data fails closed
Fields that look like PHI are flagged and excluded from AI prompts, charts and exports until a person decides otherwise. Treating a field as less sensitive is a justified request an owner approves, not a silent edit.
Nothing is overwritten
Raw uploads are preserved exactly as received. Change something upstream and everything derived from it is marked stale rather than quietly rewritten.
Start with a question and a spreadsheet.
No plan to choose, no card to enter. Bring your own data or generate a synthetic population to try the loop first.