Analytics & Insights - AI Engine
A clearer picture of your data.
A more informed next move.
Drop your datasets here
One file, multiple files, or an entire ZIP folder.
Size limit per table
100% of your data is analysed, never a sample. No limit on file size in MB.
Rows allowed = 80 million cells ÷ number of columns
Larger tables are not loaded. Device memory may require smaller files even below these limits. Learning methods have separate full-record limits: 3,000 records for Machine Learning, 150 for hierarchical grouping, and 60,000 eligible records for prediction and what-if modelling. Beyond a method’s limit it is not run, never sampled.
Set sharing to "Anyone with the link". The file passes straight to your browser and is never stored.
The whole picture, not just the numbers.
Descriptive
What happened?
Patterns, trends & comparisons
Diagnostic
What may explain it?
Relationships & possible drivers
Predictive
What could come next?
Validated estimates & uncertainty
Prescriptive
What should we consider?
Evidence-backed next steps
Supported formats & analysis limitations
CSV, TSV and delimited text (TXT, PSV, DAT); Excel (XLSX, XLS, XLSM, XLSB) and OpenDocument spreadsheets; Parquet; Arrow and Feather; JSON, JSONL, NDJSON and GeoJSON properties; XML records; SQLite database files and self-contained SQLite-compatible SQL dumps (tables and views; not arbitrary MySQL/PostgreSQL scripts or live database connections); PDF page text; Word DOCX tables and paragraph text; GZIP-compressed files; ZIP folders containing any of these. Single or multiple files. PDF imports readable text, not confirmed numeric table columns; scanned PDFs need external OCR. Word supports DOCX, not legacy DOC. Word tables use the first row as headers; verify merged cells and layout in Original data. For reliable numeric analysis of PDF tables, export to CSV or Excel. SAS, SPSS, Stata, Avro, ORC and images are not analyzed yet; export them to CSV, Excel or Parquet first.
Big files: there is no fixed file-size rejection for local uploads, but capacity is not unlimited and depends on your device. CSV and delimited text files over 50 MB are streamed in chunks and Parquet files are read in pieces, but the complete parsed table and its analysis still need to fit in browser memory. Every row is analyzed in full, up to about 80 million cells per table (for example 10 million rows × 8 columns, or 4 million rows × 20 columns). Larger tables are not loaded or sampled; the app explains how to split them. Progress is shown while files are read and analyzed. PDF, Word, SQL, spreadsheets, JSON, Arrow/Feather, other non-streamed formats and ZIP contents are loaded whole and must fit in browser memory; the practical limit depends on your device. No method silently samples your records. Nested ZIPs have a depth limit.
Original source files are not overwritten. Empty cells become missing values; numbers and dates are interpreted read-only. Spreadsheet dates are represented as ISO dates; JSON objects may be flattened. Model inputs may be imputed for modeling only. Files, results and context are session-only and are lost on refresh; download your report before leaving.
Column meanings and default outcomes are inferred from names and values. Confirm units and business context. Numeric outcomes use inferred sum or average aggregation; flag outcomes use rates. Regression/classification use a reproducible random holdout, not a time-based future test. Forecasts use a chronological holdout; intervals are approximate. Not every dataset supports a reliable prediction.