See the shape of your data. Scatter Lab is a fully client-side workbench for exploring tabular datasets as interactive 2D/3D scatter plots — built for survey and questionnaire research, useful for any table of numbers.
Live: scatter-lab.vercel.app
All computation runs locally in the browser. Parsing, PCA, clustering, statistics, and persistence all happen in the tab, and your dataset is never uploaded to a server. There is no backend; the app deploys as a static site.
The one exception is the optional AI assistant. If you connect an API key, column names and aggregate summaries — never raw rows — are sent to whichever provider that key belongs to. See The assistant below; leave it disconnected and nothing leaves the tab at all.
The built-in demo is Fisher's iris data, taken from the UCI Machine Learning
Repository's bezdekIris.data (doi.org/10.24432/C56C76).
The file matters more than it sounds. The widely-mirrored Kaggle copy carries
UCI's documented errata in rows 35 and 38, which is enough to move the first
principal component from 72.96% to 72.77% — a difference that reads as a bug in
the eigensolver rather than a difference in the data. bezdekIris.data is the
corrected version, and the app's PCA now agrees with R's prcomp to four
decimal places. Provenance, both checksums and a one-line reproduction command
are in
frontend/public/demo/iris.SOURCE.md.
- Drop in CSV, XLSX, or Parquet — every column is profiled automatically (type, range, missing values, live mini-histograms) with one-click X/Y/Z/color assignment from the Variables panel. Smart first-view defaults avoid explicit ID columns and prefer a small non-boolean categorical grouping for colour. Multi-sheet workbooks offer a sheet picker, and a sheet with no data rows is skipped rather than reported as empty.
- It tells you what it did to your file — ragged rows, unterminated quotes, duplicate or blank headers, numbers written with decimal commas or thousands separators, and dates read as text are all reported rather than silently absorbed. Nothing is auto-corrected.
- It finds missing-value codes, and says how sure it is — SPSS and Qualtrics
write "don't know" and "refused" as ordinary numbers (
-99,-999,9999, or a 9 on a 1–7 item), and a CSV carries no sign of it. Three rules look for a wrong sign, a value far outside its column, and a hole in a short scale; each finding is labelled certain, likely or possible. A cross-column check reports whether the same respondents carry the code across a battery of items — the signature of a Don't Know block — and says how much more often than chance. You can also declare the codes your survey used, which is matched literally. - Blanking them is per column, and it shows its work — the same 9 can be a code in a Likert item and a real age in the next column, so the choice is yours per column, with boxes pre-ticked only where the evidence is strong. Every recode reports cells blanked, n before and after, and the shift in mean and sd, and asserts that everything it did not target is unchanged.
- In-app PCA — pick variables, pick components, run: a browser-side
eigensolver produces scores, loadings, and a scree chart (correlation- or
covariance-based). Missing values can be median-imputed, reconstructed by
regularized iterative PCA (the missMDA
imputePCAmethod), or dropped complete-case; every run reports exactly what it filled or dropped, and in which variables. A precomputed components/loadings file works too, and now says how much of itself it actually matched - Clustering — DBSCAN and K-Means (k-means++, deterministic) with live parameter sliders, per-cluster composition breakdowns, and diagnostics (silhouette-by-k, k-distance percentiles for eps). Save composition heatmaps as PNGs in Viridis, Inferno, or Greens, with a 0–100% colour-scale legend.
- Compare views — pin up to 4 views in a tiled grid. A pin keeps the data and framing it was taken with while the live view moves on, but stays fully interactive: rotate, zoom or double-click any pane to reset that pane alone. Transfer columns (e.g. cluster labels) between datasets by row order or key match, with alignment guards
- Exports — PNG, rotating GIF, or a fully self-contained interactive HTML file that works offline
- Workspaces — sessions persist locally (IndexedDB) and export/import as shareable files
- Two themes — Bauhaus and Terminal
Load demo on the opening screen is the short way in: one click loads the Iris data, opens the panel, and starts the walkthrough. Opening the assistant panel directly offers the same choice as a menu.
The guided walkthrough is a scripted, click-through tour of the Iris demo — eight steps, one button at a time, no API key and no network. It drives the real workbench as it goes (loads the data, runs the PCA, clusters the scores, pins a comparison) and points at each control with a highlight that holds until you move on. Uploading is paused while it runs and it never starts a download. If a dataset of yours is already loaded it says what it is about to change and offers to save a workspace first, and the state you had is one click away afterwards.
The message box is disabled throughout — the buttons take you through, Exit demo sits on the end of that row, and a step list below the transcript shows where you are and what is left.
The steps live in frontend/src/lib/walkthrough.ts as data, so the tour is the
same every time and a broken step is a failing test rather than a dead end.
An in-app AI copilot that drives the workbench through tool calls: assign axes, run PCA and clustering, read cluster compositions, configure and save their heatmaps, scatter exports, or a rotating 3D GIF, compute correlations and group comparisons, tour your own data — and literally point at the interface with an ephemeral highlight while explaining it. Interpretation questions are grounded in a curated, citation-backed methods reference that ships with the app. Asked for a demo, it offers the scripted walkthrough above rather than improvising one.
- One-click OpenRouter connect (OAuth PKCE) or paste any key; any OpenAI-compatible endpoint works, including local runtimes (Ollama) for a fully offline assistant. Keys live in your browser only
- Privacy contract: requests carry column names and aggregate statistics — never raw data rows. A categorical value is only named if it covers at least five rows, so an email address, a participant name or a free-text answer is counted but never sent
- Undo for anything the assistant changes; optional thumbs-up/down feedback per reply (stored write-only, used to improve the assistant)
- Dockable right/bottom or floating, resizable, markdown-rendering chat
Because users never see the code, the choices the app makes on their behalf are
written into the interface: small (i) markers next to the controls they
apply to, and an About page reachable from the header. Both read from one
source (src/lib/disclosures.ts), so a tooltip, the About page and the
assistant's own reference cannot drift apart. They cover the things that change
an answer — that K-Means here is the deterministic variant, that clustering runs
on the plotted axes only, what standardizing does, and which of the two
conventions the reported loadings follow.
cd frontend
npm install
npm run dev # http://localhost:3000
npx vitest run # unit tests
npx tsc --noEmit # types
npx next build # production buildCI runs all four on every push and pull request, plus a lint gate that allows the existing warning count to fall but never rise.
Feedback storage is env-gated: copy .env.local.example to .env.local and
fill the Supabase values to enable it (unset = feedback UI hidden). Pushes to
main deploy to production via Vercel; tests run in GitHub Actions.
Built with Next.js, Plotly (WebGL), Tailwind CSS, PapaParse, SheetJS, and hyparquet. See HANDOFF.md for architecture and project state.
The original FastAPI/pandas backend (removed in the client-side migration) lives in git history for reference.