Drop in your project, thesis or business spreadsheet — ATARS cleans it, runs the statistics, charts it, trains and explains ML models, and writes the finished report in minutes, as a private desktop app on your own computer. No code, no setup, no cloud.
Installs like any desktop app · runs fully offline · your data never leaves your computer
✓ Now live on the Microsoft Store · direct installer · pip install atars for the terminal
// Built by a student. Designed for every learner.
Most people get stuck between a messy spreadsheet and a meaningful answer. ATARS closes that gap with intelligent automation, so domain experts don't need to be data scientists.
Install once and go. Cleaning, SQL, charts, statistics, modelling, forecasting and reporting — a guided workflow in a real desktop app, no Python, no servers.
ATARS reads your dataset, infers the problem type, recommends a workflow, and suggests the right charts and models automatically.
ATARS runs fully on your own computer — files are processed locally and never uploaded anywhere. AI calls happen only if you add your own key, straight to your chosen provider.
Healthcare, climate, finance, agriculture, retail, research — any structured tabular dataset, any domain.
Export polished PDF, Word, PowerPoint, Excel and Markdown reports — plus a runnable Jupyter notebook — in one click.
Built by a student to be genuinely affordable and accessible for every learner, researcher and small team.
The tools you'd normally need a database, a notebook and a BI seat for — all built in, no servers to wire up.
Query your data in full SQL — joins, windows, CTEs — powered by DuckDB. Connected to a live database? Push heavy SQL to the database itself, so tables far bigger than memory still aggregate in seconds.
Cohort & retention heatmaps, RFM customer segments, and time-series decomposition — one click each. None of these exist as buttons in Excel.
Stack or join several files, then compare every branch side by side — with real significance tests (ANOVA, Kruskal-Wallis, pairwise t-tests) to tell signal from noise.
Named after the ancient Indian mathematician. Ask anything and it runs the analysis your question needs, writes and runs the code, and shows every step. It reads your whole dataset — every column's stats, correlations and range — and every figure it writes is checked against the numbers ATARS actually computed, with a visible grounding score. No hallucinated statistics.
Ask about live trends, prices or regulations and Aryabhata searches the web in real time — answering with insights and source links as proof, ready to fold into your report.
Save your whole setup — cleaning rules, validation checks and pinned charts — as a recipe file. Next month's file arrives, load the recipe, and the same analysis rebuilds in one click.
Every figure carries its working — the column, the filter, how many rows were left out and why, and the arithmetic. When you export, each one is re-checked against your data and marked as holding or moved, so a number in a report can prove itself months later.
Define a business term once — "Revenue = sum of Sales, excluding returns" — and every chart and report uses that same definition. Two people can no longer present different numbers for the same word.
Add policies, contracts or past reports and ask about them alongside your data. ATARS cites the passage it used, and remembers your documents between sessions.
Aryabhata doesn't improvise its method. It follows 18 notes on how careful analysts work, written for ATARS and drawing on 49 published works. For each question it looks up the notes that apply — on your own computer — so an answer knows when it may say "causes", when a sample is too small to trust, and when a model looks too good to be true. The numbers still come only from what ATARS computed.
Choices made after seeing the data make results look surer than they are. A pattern found while browsing is reported as worth testing on new data — not as a finding.
Draws on Gelman & Loken (2013); Simmons, Nelson & Simonsohn (2011)
A correlation licenses no causal claim on its own. What licenses one is a design, or an argument about how the data came to be.
Draws on Pearl (2009); Hernán & Robins (2020)
Information that won't exist when you actually predict is the most common reason a model scores brilliantly and then fails in use.
Draws on Kaufman, Rosset & Perlich (2011); Kapoor & Narayanan (2023)
Why a value is missing decides what you may do about it. Some gaps no imputation can fix — they have to be modelled, or stated as a limitation.
Draws on Rubin (1976); Little & Rubin (2019)
Common fairness definitions can't all hold at once, so choosing one is a decision for people, not the model — and a trend can reverse inside every group.
Draws on Kleinberg, Mullainathan & Raghavan (2017); Bickel, Hammel & O'Connell (1975)
Small samples don't just miss real effects — the results that do reach significance are inflated, sometimes several times over, and can have the wrong sign.
Draws on Cohen (1988); Gelman & Carlin (2014)
These notes are ATARS's own summaries. The authors wrote the original works and are not affiliated with ATARS. The notes guide how the AI reasons; they don't make it infallible, so check any answer a decision depends on. All 18 notes and their sources →
ATARS compares its own results with pandas, NumPy, scikit-learn and SciPy on data where the right answer is known in advance. 14 of 14 correctness checks pass on the current code.
Mean, median and grouped totals agree with pandas, and ATARS's fast DuckDB engine agrees with pandas across six aggregations — the only differences are floating-point rounding.
Given data built with a known answer, ATARS removes the duplicates, recovers the clusters, flags the planted outliers, catches a column that leaks the answer, and ranks real signal above noise.
More than 900 automated tests run against the engine, including checks that fail if a core calculation silently changes its result.
These checks cover ATARS's core calculations, not every result it can produce, and they are run by the developer rather than an independent auditor. Matching the reference libraries shows the arithmetic is right; choosing the right method for your data is still a judgement call. See every check and its result →
Production modules under one roof — each one designed to remove a step that normally needs a specialist.
Preview, profile and edit in real time. Multi-sheet Excel? Pick a sheet or combine them all. Report-title junk above the table? The real header row is detected automatically.
Infers the target, features, dates and problem type on load, then recommends the workflow and models to run.
A readiness score from completeness, consistency, duplication, outliers and balance — plus leakage & bias flags.
Fixes real-world mess: currency & crore/lakh numbers, European decimals, duplicate category spellings ("Male"/"male"/"MALE" → one), missing values, outliers — every step logged in an audit trail.
Generate date, lag, rolling-window and interaction features to strengthen every model and forecast.
15 interactive chart types with slicers and pinnable dashboards — plus optional AI chart suggestions.
t-test, z-test, ANOVA, chi-square, normality, correlation, non-parametric tests and assumption diagnostics.
Trend & seasonal forecasts with confidence bands and a holdout backtest, so you can trust the numbers.
Compare a reference dataset against today's to catch distribution shift before it breaks your models.
Data, cleaning, charts and notes live in one .atars file — with several sheets inside it — that you save, reopen and send like any document.
Comments, reviews and case notes are grouped into themes and lined up against your numbers — computed on your own computer, over every row.
Install a local model from inside ATARS for AI that never leaves your machine — or bring a free Groq or Gemini key. Work is sized to your computer's cores and memory.
In the download above — install v5.4.7 and these are already there. Full list in the changelog →
A column that is mostly empty keeps its gaps instead of being filled with one repeated value, and free-text notes are never filled with words nobody wrote. The cleaning log names every column it left untouched, and why.
When a provider retires a model, ATARS finds one your key can still use, switches to it and tells you — and your questions and answers are saved inside your .atars file.
Download buttons no longer rebuild their files on every click — on a 10,000-row table, about 0.7 s down to 0.02 s — and switching sections no longer leaves the last screen looking frozen.
ATARS organises every dataset into a natural left-to-right flow — Understand, Prepare, Explore, Model, Report. Stop, branch or export at any step.
Load any file or database; ATARS auto-detects structure and problem type.
Score data quality, then catch missing data, duplicates and outliers.
Clean, combine files, engineer features, or reshape with SQL.
Charts, dashboards, statistical tests and business analytics.
AutoML ranks a leaderboard; train, validate and cross-validate.
SHAP global & local explanations and driver analysis.
What-if predictions, sensitivity and root-cause analysis.
Export a narrated PDF / Word / PPT / Excel / notebook.
Drop two or more files to combine them — stack rows or join on a key.
Full tables load with no silent row caps — what you connect is what you analyse.
Train, compare and trust your models — with the reasoning made visible.
Trains multiple models and ranks a leaderboard so the best one is chosen for you — with leakage-safe cross-validation, a live training progress bar, and a Fast / Full-accuracy switch to trade speed for precision on big data.
Regression, classification and clustering with train / validate / cross-validate, a performance view, and single & batch prediction.
Global & local explanations, feature importance and contribution, and per-prediction breakdowns — see why, not just what.
Trend, seasonal and growth forecasts with confidence intervals and a holdout backtest for any time-indexed series.
What-if predictions, sensitivity testing and root-cause analysis — change a variable and see the impact instantly.
Compare datasets or branches with real hypothesis tests — ANOVA, Kruskal-Wallis and pairwise t-tests — to know a difference is real.
ATARS doesn't stop at a chart on a screen. Generate a complete, narrated analysis — your charts, stats, model results and audit trail — as a document, ready for a class submission, a research appendix or a leadership review. Add your own AI key and every export gains an executive narrative, interpretation, risk callouts and recommendations — plus an AI verification section showing how many of the narrative's figures were checked against your computed data, and an optional web research appendix with cited sources.
Honest examples of what ATARS does end-to-end — bring your own data and follow the same path.
Run descriptive stats, normality and significance tests on your study data, then export a clean statistical appendix you can attach to a thesis.
CSV → Statistics → PDF
Load historical sales, let ATARS detect seasonality and trend, and produce a forward forecast with confidence bands for a planning review.
XLSX → Forecasting → PPTX
Connect your orders table, query it in SQL, then build RFM segments and retention cohorts to see who to keep and who's churning.
SQL → RFM & cohorts → Excel
Open, proven Python data tools — compiled into a native desktop app (machine code), so nothing needs installing separately and nothing is locked behind a vendor.
A real desktop app, now live on the Microsoft Store. Install once, analyse forever — your data never leaves your computer.
Windows 10 (1809+) / 11 · 64-bit · ~505 MB download
Or open the Store app on your PC and search ATARS — publisher Priyanshu Kumar, product ID 9NFRBPL4T2M8.
Windows 10 / 11 · 64-bit · ~350 MB download · ~1.2 GB installed
Heads up: Windows will show a "Windows protected your PC" notice — normal for an app that is not code-signed, not a virus warning. Click More info → Run anyway, or take the Store version instead. Why this appears →
Verify your download — SHA-256:1871ec7b4d2eb6c9bb44f80dd756cb8bf1374786b71bd7131eee2bb2c2b8816b (v5.4.7)
Windows · macOS · Linux · Python 3.11+
pip install atarsatars for a full-screen window with clickable tabsNot as powerful as the app. It is a deliberately limited edition: no AI analyst, no Advanced Statistics, 17 chart types instead of 19, and no Word / PowerPoint / notebook exports. A terminal also cannot draw a treemap, sunburst or bubble chart honestly. For the full product, use the desktop app.
🐧 Linux and 🍎 macOS desktop builds are planned. Until then, the terminal edition above runs the same engine on both. Tell me when the desktop build is ready →
B.Tech Computer Science · Gurugram University (2024–2028)
ATARS started as a question: why should advanced analytics need a data-science degree? It's now a universal platform on a mission to make rigorous, explainable data work available to students, researchers and small teams everywhere — proudly built in India.
.exe if the Store is blocked or unavailable on your PC, as it often is on school and work machines. The terminal edition (pip install atars) is deliberately smaller: same engine, but no AI analyst, no Advanced Statistics, 17 chart types instead of 19, and no Word / PowerPoint / notebook exports. It exists for terminals, servers, and for macOS and Linux users while the desktop builds are in progress — not as a replacement for the app..atars analyses) and databases & warehouses (SQLite, MySQL, PostgreSQL, Amazon Redshift, Snowflake, Google BigQuery). You can also drop several files and combine them by stacking rows or joining on a key.Download ATARS, install it in a minute, and run a full analysis on your own computer — free, private, no sign-up. Three ways to get it: the Microsoft Store (now live), the direct installer, or pip install atars for the terminal.