NiftyStats: a statistics dashboard that never sees your data

Drop a spreadsheet on the page and get descriptive statistics, a correlation matrix, regression, a time-series trend and a clustering pass back, each one sitting next to a sentence in plain English that says what it means. The maths is real Python. It runs inside the tab.

Live demo: https://igorcsis.github.io/niftystats/
Stack: React and Vite for the page, CPython compiled to WebAssembly for the statistics. No account, no API key, no upload.

The problem it solves

A small business owner exports a spreadsheet and gets stuck at the same place every time. The numbers are all there. What is missing is somebody to say which of them matter. The usual answers are a pivot table that answers a question you already knew to ask, or a paid tool that wants the file uploaded to somebody else's server first.

NiftyStats runs the analysis a statistician would run first, then writes the answer out in a sentence. The file stays on the machine it started on.

Where the data goes

There is no backend to send a file to. The page is a static bundle on GitHub Pages, and the statistics engine is CPython compiled to WebAssembly, running in the same tab as the file picker. Nothing in src/ calls out anywhere except to fetch the engine itself.

Two third party requests do happen, and it would be dishonest to draw them out of the picture: the Pyodide CDN, and Google Fonts for Inter and JetBrains Mono. They see an IP address. They do not see a spreadsheet.

The one-time wait

Running real Python in a browser means downloading a real Python first. That is a one-time cost and the page says so while it happens, step by step, rather than parking a spinner on the screen and hoping.

public/sw.js caches exactly the cdn.jsdelivr.net/pyodide/ URLs and nothing else, so a returning visitor pays for none of it again. The app's own assets are left to the browser's HTTP cache, where Vite's fingerprinting already handles them.

What you get from one file

Pass	What it actually computes	Module
Descriptive	Per column: mean, median, standard deviation, MAD, percentiles at 1, 5, 25, 50, 75, 95 and 99. Outliers counted two ways, by Tukey's IQR fence and by a modified Z-score at 3.5. Shapiro-Wilk normality up to 5,000 rows. A histogram capped at 40 bins	src/python/descriptive.py
Relational	Pearson and Spearman matrices with a p-value per pair, and a flag on any pair where the two disagree by more than 0.20, which is usually a curve rather than a line	src/python/relational.py
Linear regression	Ordinary least squares with standardized coefficients, t-based p-values, and a multicollinearity flag on any feature whose VIF passes 10. Needs 20 usable rows	src/python/relational.py
Logistic regression	Unregularized, on a boolean target, with Wald-test p-values from the inverse Fisher information. Refuses to fit under 20 usable rows or 5 in the minority class	src/python/relational.py
Time series	A scipy.stats.linregress trend per numeric column, needing at least 10 points, forecasting ahead by a quarter of the series up to 12 points, with a 95% band from the residual standard error	src/python/timeseries.py
Clustering	K-means for k from 2 to 8, picking k by silhouette score and flagging anything under 0.25 as a weak split. PCA to two dimensions for the scatter plot only, never for the clustering itself	src/python/clustering.py

Every one of those numbers then goes through a narrative layer that turns it into a sentence. The narratives are templates keyed off thresholds, so the same file always produces the same words.

What it deliberately does not do

• It does not upload your file. There is nowhere for it to go. See the diagram above.
• It does not call a language model. The plain-English layer is src/lib/narratives/, a set of templates driven by the numbers. Same input, same sentence, every time, offline.
• It does not remember anything. No account, no localStorage, no database. Close the tab and the data is gone with it.
• It does not handle very large files. PapaParse in worker mode is comfortable to roughly a hundred thousand rows. Past that the engine would need a chunked pass it does not have yet.
• It does not replace a statistician on a decision that matters. It runs the first pass and tells you where to look.

Running it locally

pnpm install
pnpm dev        # http://localhost:5173
pnpm lint       # eslint over the whole tree
pnpm build      # production bundle in ./dist
pnpm preview    # serve that bundle locally

The Python modules are audited separately, because they follow the Appendix A conventions rather than the surrounding TypeScript style:

python tools/appendix_a_audit.py src/python/*.py

Both the lint and the audit run in CI on every push.

Layout

niftystats/
├── index.html                        the shell and the font preconnects
├── src/
│   ├── main.tsx                      mounts React, registers the service worker
│   ├── App.tsx                       the single route
│   ├── pages/Landing.tsx             the upload screen and the dashboard below it
│   ├── components/
│   │   ├── upload/                   drop zone, sample buttons, engine status
│   │   ├── dashboard/                one section per analysis pass
│   │   ├── charts/                   the Plotly wrappers and their shared theme
│   │   └── layout/                   header and footer
│   ├── lib/
│   │   ├── pyodide/                  the engine bootstrap, and its status feed
│   │   ├── narratives/               the plain English layer, one file per pass
│   │   ├── pdf/                      the export, html-to-image into jsPDF
│   │   └── csv.ts                    hands a file to the worker and waits
│   ├── python/                       the four analysis modules, run inside Pyodide
│   ├── workers/                      PapaParse or SheetJS, and the type sniff
│   └── types/                        the contract between the two languages
├── public/
│   ├── sw.js                         caches the Pyodide CDN payload, nothing else
│   └── samples/                      three fictional CSVs, 20 to 30 rows each
├── tools/appendix_a_audit.py         the Python style audit CI runs
└── assets/                           the diagrams in this README

ARCHITECTURE.md has the longer version: the data flow, the boundaries between the layers, and what is deliberately left for later.

The sample data

Three fictional CSVs ship with the app, reachable from the buttons under the drop zone:

File	Rows	Shape
sales-sample.csv	20	A date, marketing spend, revenue, customers, channel
marketing-sample.csv	20	Campaign, channel, impressions, clicks, conversions, spend, start date, active
customers-sample.csv	30	Customer id, signup date, age, plan, monthly revenue, support tickets, churned

The companies, the customers and the numbers are all invented. They exist to exercise the analysis paths, not to describe anybody.

Deploying your own copy

Push to main and .github/workflows/deploy.yml lints, audits, builds and publishes dist/ to GitHub Pages. Vite is configured with base: '/niftystats/', so asset paths resolve under the Pages subpath.

If you fork this, there is one manual step: Settings, then Pages, then Source, then "GitHub Actions".

One thing to keep in sync if you upgrade: src/lib/pyodide/client.ts hard codes the CDN URL for a specific Pyodide version, and package.json pins the matching JS bindings exactly. Bump both together or the wasm runtime and the bindings will disagree.

License

MIT.