Scatter Lab

The problem

Exploratory analysis benefits from quick visual iteration, but research datasets can be sensitive. Lightweight visual tools often ask users to hand their data to a server before any of that work can begin. Scatter Lab is designed around the idea that analysis should happen where the data already is.

The interaction

Drop in a CSV, XLSX, or Parquet file. Scatter Lab profiles each column, chooses a sensible first view, and lets the researcher map variables to X, Y, Z, color, and marker shape. Run PCA in the browser and inspect variance explained, score axes, and loadings; explore DBSCAN or deterministic k-means clustering; compare cluster composition by group with normalized heatmaps; and pin a live view alongside up to three comparison views.

Exports include a PNG, rotating 3D GIF, self-contained interactive HTML file, or the active derived dataset.

An optional analysis copilot

An opt-in assistant can operate Scatter Lab through explicit interface tools: assigning plot axes and encodings, running PCA or clustering, inspecting composition, configuring heatmaps, pinning views, controlling 3D rotation, and preparing PNG, GIF, HTML, or CSV exports. During a guided walkthrough, it can highlight a relevant control, explain a default, run the next analysis step, and pause for the researcher’s direction. It can also retrieve a curated, citation-backed methods reference before interpreting PCA, clustering, or statistical outputs.

The copilot is deliberately bounded: users connect their own OpenRouter or other OpenAI-compatible key, and a local runtime can keep it fully offline. It receives column names, data profiles, and aggregate results—not raw dataset rows; asks before downloads; and every analysis or view change can be undone in one action. It is a guided interface layer over local analysis, not a black-box service that takes possession of research data.

Local-first by design

Scatter Lab is a static web application. Parsing, profiling, PCA, clustering, statistics, export rendering, and workspace persistence all run in the browser. A dataset is never uploaded as part of normal use.

System design

Built with Next.js, React, TypeScript, Tailwind, Plotly WebGL, and IndexedDB, Scatter Lab pairs a columnar table with browser-native parsing, analytics, comparison, and export generation. IndexedDB stores workspaces on the device. The key design choice is keeping the full loop—data inspection, analysis, comparison, and export—fast, inspectable, and local.