Skip to content
Celeus

Getting started

This walkthrough takes one dataset from upload to a downloaded result. It uses a small made-up file, trial.csv, with one row per participant and three columns: arm (drug or placebo), age, and sbp_change (the change in systolic blood pressure). Any dataset of your own works the same way.

Sign in at app.celeus.ai with your Google account, then open Upload. Drop the file onto the upload area or click to choose it.

  • CSV, TSV and Excel files work on every plan. SPSS, Stata and SAS files need a paid plan.
  • If the column names are not on the first row (a title line or blank rows above them), Celeus shows you how it will read the file and asks you to confirm before anything is saved.
  • The file you upload becomes version 1 of the dataset. Later changes, such as cleaning or de-identifying, create new versions and never overwrite it.

2. Check the profile and the identifier scan

Section titled “2. Check the profile and the identifier scan”

After upload you land on the dataset page. The profile lists every column with its type, its role (for example continuous, binary or categorical), the share of missing values, and a short summary.

Every upload is scanned for identifiers such as names, dates, record numbers and free text. A flagged column is marked needs review and is excluded from the AI until you decide otherwise. Review each one: you can keep it, drop it, pseudonymize it, reduce dates to the year, cap ages over 89 or truncate ZIP codes. Applying de-identification writes a new dataset version.

The Data check panel lists data-quality findings, such as duplicate rows or spelling variants of the same category, with a shortcut to fix them in Prepare data.

Optionally, open Research context for the AI and describe the study in a sentence or two. The AI uses this text when it suggests methods and writes interpretations. Never put row values or patient details in it.

The dataset page offers several ways in. The three most common:

  • Suggest analyses ranks candidate analyses from your columns. The ranking is computed from the dataset profile, with no AI involved, and each candidate is graded from assumption checks on your data: Recommended, Suitable, Use with caution or Not appropriate. You can also ask the AI for an advisory recommendation beside the ranking.
  • Configure manually lets you pick any test yourself.
  • Chat with your data lets you describe what you want in plain language; the assistant proposes an analysis and you decide whether to run it. See Use chat and Ask AI.

For trial.csv, a suggestion such as comparing sbp_change between the two arms is a natural first choice.

The configure screen shows the chosen test and one slot per role it needs (outcome, group, predictors and so on). Map your columns into those slots. Then check:

  • Missing-data strategy - how rows with missing values are handled. See Handle missing data.
  • Analysis population (optional) - restrict the analysis to a subgroup without changing your data.
  • Research question (optional) - recorded in the package and used by the AI interpretation.
  • AI interpretation and manuscript draft - optional prose written by the AI from the computed results. The statistics are the same with or without them.

Choose Run analysis. The results page opens and fills in when the run finishes.

The results page shows the headline result (n, estimate, confidence interval, p-value, effect size), the assumption checks, any warning flags, a plain-language summary computed without AI, and the AI interpretation if you asked for one. Reading your results explains each part.

At the top, Verify before accepting asks you to check what was analyzed, the numbers and the flags. When you are satisfied, choose Confirm these results (or Confirm results and unlock download at the bottom of the page). The confirmation is saved with the run, with who confirmed it and when.

Confirming unlocks Download package (.zip): the reproducible package with the analysis plan, the results and tables, the exact script, the seed, the package versions and the audit log. See Export results and the reproducible package for this and the other ways to take results out, and The reproducible package for what is inside.