Every figure traced to a computation
Ask questions of your spreadsheet. Get answers you can check.
Upload a CSV or Excel file and ask in plain language. An AI model plans the analysis, deterministic tools compute every number, and each figure in the report links to the calculation behind it.
- CSV, TSV, XLSX and XLS
- Free to use
- Charts and exports
Total revenue – North America
4,851,042.44Computed inE1
E1Calculating sum of revenue, mean of unit_price by regionaggregate
| region | rows | sum_revenue |
|---|---|---|
| North America | 972 | 4,851,042.44 |
| Europe | 688 | 3,266,861.56 |
| APAC | 504 | 2,384,029.35 |
| LATAM | 236 | 1,002,437.59 |
All 16 figures verified against computed results
How it works
The model plans and writes. Code does the maths.
Language models are good at deciding what to look at and explaining what was found. They are unreliable at arithmetic. So the two jobs are kept apart.
- Step 1
Upload a dataset
A CSV, TSV, XLSX or XLS file. It is validated, parsed and profiled: column types, missing values, duplicates and a quality score.
- Step 2
Ask a question
In plain language. The AI model reads the column names and summary statistics and plans which computations to run.
- Step 3
Tools compute the numbers
Aggregations, time series, correlations and other analyses run as deterministic Python and pandas code on your data. You watch each step as it happens.
- Step 4
Read a verified report
The model writes the report from the computed results. Every figure in it is then checked against those results, and any that cannot be traced is flagged.
What you get
A report you can audit
Not a chat transcript: a structured report with metrics, findings, charts, recommendations and the full method behind them.
Computed, not generated
Numbers in a report come from deterministic Python and pandas tools. The language model plans and writes; it does not do the arithmetic.
Every figure verified
Each figure in the summary, metrics and findings is compared with the computed results. Unverified figures are listed, not hidden.
Provenance on every finding
Metrics and findings cite evidence records. Open one to see the tool that ran, its arguments and the table it returned.
Automatic data profiling
On upload you get column kinds, missing-value rates, distinct counts, summary statistics and a list of data-quality issues.
Follow-up questions
Keep asking about the same dataset in one conversation. Each answer is a new report with its own evidence.
Exports
Download a report as Markdown, PDF or Excel, and any result table as CSV.
Try it on your own data
Create an account, upload a file or use the sample dataset, and ask your first question. An analysis usually takes one to two minutes.