Quantilo
Open the app

Drop a file. Know what’s wrong, what drives it, what comes next.

Cleaning, explained models and forecasts with nothing to configure. Every automatic choice is a coded rule that shows its reason.

Open the app See pricing
house_sales.csv 21,613 rows · 21 columns
  1. 1.1 sColumn types21 columns read: 12 continuous, 4 discrete, 2 dates, 1 identifier, 1 yes/no, 1 postal code. 3 to confirm.
  2. 2.4 sFixed for youdate is text such as 20141013T000000: stored as dates. Applied, rule right on 99% or more of cases
  3. 2.4 sYour callyr_renovated is 0 on 20,699 rows (95.8%): it means never, not the year 0. Waits for your tick
  4. 5.0 sFirst modelA linear regression predicts price with R² 0.69 on rows it never saw.
  5. 18 sWhat drives the pricegrade, zipcode and sqft_above: the score drops most when each one is shuffled.
  6. 44 sNext 30 daysAverage price per day, forecast by ETS: 27% more accurate than a seasonal naive forecast over 5 back-tests.

Real output of the engine on the King County house sales sample (CC0). Times measured with the desktop engine on a laptop; the browser version first loads for about a minute. Try this sample in the app.

From a file you have never seen to something you can say about it.

  1. 01

    Drop a file

    CSV, Parquet or Excel, or a folder of files with the same columns. Encodings, separators and title rows above a table are read for you.

  2. 02

    Read the verdicts

    Each column gets its type and role with the reason, the fixes it needs, and the right chart and test. Nothing to set first.

  3. 03

    Go as deep as you need

    Train and explain a model, forecast a series, build KPIs. Save the work as a recipe and replay it on next month’s file.

The same file, in the app.

Screens of the free browser version on the 21,613 house sales above.

The Clean page: 4 fixes found. One is applied automatically on 21,613 cells (dates stored as text); three need a decision, such as 13 rows with 0 bedrooms and 0 read as never in yr_renovated, each with its reason.
What’s wrong. One fix applied on its own, three left to you, each with its reason and the cells it changes. The original file is never modified.
The Correlations page on the house sales: all 141 significant links stay significant after correcting for 153 tests (Benjamini-Hochberg); strongest first, such as sqft_above rising with sqft_living (r = 0.88), and the skewed columns.
What moves together. Every pair of columns with the measure that suits their types. Here, 141 significant links all hold after correcting for 153 tests.
The Forecast section of the time series page: the average price per day with ETS forecasting the next 30 days and its 80% and 95% intervals, 27% better than the seasonal naive reference over 5 rolling origins.
What comes next. Pick a date and a number: the models are compared on back-tests, and the best one forecasts with its intervals. This one ran in the browser in about 10 seconds.

Ten layers. Each one starts from what the last one found.

The types choose the tests, the cleaning feeds the models, the models feed the decisions. You can stop at any layer.

  1. 01

    Read

    CSV, Parquet, Excel, JSON, logs and folders of files. The encoding is read from the bytes; tables inside a laid-out sheet are found.

  2. 02

    Type

    Numbers, categories, dates, identifiers, free text, and roles such as amount, surface or postal code, each with its reason.

  3. 03

    Clean

    78 cleaning rules. The 14 measured right on 99% or more of cases apply themselves; the others wait for your tick.

  4. 04

    Describe

    Every variable at a glance, then one in depth: normality tests, law fitting, outlier rules.

  5. 05

    Relate

    Correlations across every pair of types, p-values corrected for multiple testing, regression with controls, structure (PCA).

  6. 06

    Group

    Segments of similar rows, and the unusual rows that stand out on several columns at once.

  7. 07

    Predict and explain

    39 models in 10 families and an Auto mode. SHAP values, effects of each column, and what would change a prediction.

  8. 08

    Forecast

    Trend, seasonality, stationarity, and 29 forecasting models compared on back-tests, pretrained AI models included.

  9. 09

    Decide

    KPIs suggested from your columns, A/B tests, sample sizes, and what other decision policies would have earned.

  10. 10

    Replay and share

    Recipes that replay the work on a new file, reports, exports, and synthetic copies without the real rows.

Counted in the code, not estimated.

machine-learning models
39
forecasting models
29
cleaning rules
78
automated tests
3,343
languages, English and French
2

Read from the source on 13 November 2026. Models: 10 families; XGBoost, LightGBM and CatBoost run in the desktop version. Forecasting: 5 pretrained models, downloaded on demand in the desktop version. Cleaning: rules and their variants, 14 of them applied automatically. Tests: the test functions of the engine, the server and the interface.

Automatic, never blind.

Quantilo decides a lot for you. Each decision is a coded rule you can read, not a guess from a language model.

  • Every choice shows its reason

    “Name suggests amount”, “only 13 distinct whole numbers: shown as categories”. Rules carry a code you can look up in the guide.

  • Measured before it acts

    A cleaning rule applies itself only when it was right on 99% or more of the cases measured on real files. The others wait for you.

  • Warnings against the usual mistakes

    Columns that look like they contain the answer are left out of a model, p-values are corrected when many are tested, panel data is split by time.

  • One click back

    Types, fixes and settings can be changed back. The original file is never modified.

The Columns page: each column with its detected type, role, confidence and the reason, such as name suggests amount for price, or only 13 distinct whole numbers for bedrooms.
Types with their reasons. Three columns are marked to check, with why; one click changes a type.

Your files stay with you, wherever it runs.

The same engine in three places. None of them uploads your files; the one optional exception, the AI assistant, is described in the questions below.

In your browserFree On your computerPro On your serverTalk to us
Install Nothing: open the page. The desktop app, on Windows, macOS or Linux. Docker, on a machine of your network.
Your data Read by the page, kept in this browser. Stays on your disk; works offline. Stays on your network.
Sizes Medium-size files. Millions of rows. Millions of rows, shared by a team.

Free in the browser. Pro in early access.

Free

For a file to understand today.

€0no account

Runs in your browser.

  • Every analysis, in your browser
  • No account: continue as guest
  • Medium-size files (the memory of a browser tab)
  • Machine learning with the scikit-learn models
  • English and French
Open the app

Pro Early access

For analysts who open files every week.

€19per month

or €190 a year · 14-day free trial

  • Everything in Free
  • The desktop app: millions of rows, offline
  • XGBoost, LightGBM, CatBoost and pretrained forecasting models
  • Recipes that run on a schedule · in development
  • PDF reports with your logo · in development
  • Connectors: SQL databases, Google Sheets · in development
  • Support by e-mail
  • Security updates included; support period stated at purchase
Join early access

Payment is not open yet. Pro is in early access: the features marked “in development” are being built, and early users say which come first. Prices in euros.

Two plans for now. For a team or a company: talk to us.

Early access opens soon.

Pro is not open yet: the first places are being prepared. In the meantime, every analysis is free in your browser, with no account. Open the app.

Questions

Does my data leave my computer?

Your files do not. In the browser version the analysis runs inside the page (Python compiled to WebAssembly): your browser reads the file and keeps it in its own storage. The desktop and server versions keep it on your disk or your network. There is no analytics, in the app or on this site, and this site sets no cookie.

What does go over the network: the browser version downloads its engine from a public CDN (jsDelivr) when it loads; if you choose “Sign in with Google”, Google handles that sign-in and the app receives your name and e-mail. And one optional exception concerns your data: the AI assistant of the desktop version, once you add your own API key, sends your questions and summaries of your data (never whole rows unless you allow it) to the provider you chose.

Does it work offline?

The desktop and server versions work without a connection. The browser version needs one to load its engine (about a minute the first time); the analyses then run in the page.

What files can I open?

CSV and other text tables, Parquet, Excel (including several tables in one sheet), JSON, JSON Lines and XML, log files, and folders of files with the same columns. Sets of images open in the computer-vision pages.

How are the automatic choices made?

By coded rules, not by a language model. Each rule has a code and a sentence: R-TS-03, for example, averages a price per period instead of summing it, “as a sum of prices means nothing”. The reason is shown next to the result, and a cleaning rule may apply itself only after it was measured right on 99% or more of the cases in real files. Every choice can be changed.

Is it free?

The browser version is free, with every analysis and no account. Pro, in early access, adds the desktop app, larger files, support, and features still in development such as scheduled recipes and reports. Payment is not open yet.

Desktop or browser: which one?

The browser version needs nothing installed and suits medium-size files. The desktop version runs the same engine natively: millions of rows, XGBoost, LightGBM and CatBoost, pretrained forecasting models, folders on your disk and the AI assistant. Today it starts from the public source code with one command; a packaged app is planned for Pro.

How large can my files be?

In the browser, the memory of one tab sets the limit (WebAssembly gives Python 2 to 4 GB, less on a phone): medium-size files. The desktop version converts a file to Parquet once and handles millions of rows; a few heavy steps use a reproducible sample and say so.

Is it in French?

Yes. The interface, the explanations and the reasons of the rules are in English and French. Choose the language in Settings.

Is it open source? Under which licence?

The source code is public on GitHub; the code published so far is under the MIT licence. The licence of the versions to come, Pro included, is being decided and will be stated here.

Who is it for?

For people who receive a file they do not know and must say something about it fast: consultants, analysts, controllers, researchers, and data scientists who want the first day of exploration done.

Open a file and see what it says.

Free, in your browser, with no account. Try it on your own file or on one of the samples.

Open the app Read the guide