Build, compare, and interrogate PCA and PLS models in your browser. ScorePilot is a point-and-click workbench for multivariate data analysis and chemometrics: load a table, clean it up, fit a model, and explore the scores, loadings, and diagnostics that explain your process - no code required.
It runs as a single, self-contained app you launch with one command. Standard chemometrics terms keep their established names throughout: scores (T), loadings (P), Hotelling's T², SPE/DModX, contributions, VIP, R²X / R²Y, Q².
Explore your data - a fast spreadsheet-style grid with a live inspector for every column (distribution, sequence, summary, transform preview), data-quality flags, and one-click roles and exclusions:
Fit a model and read it - scores with the Hotelling's T² confidence ellipse, loadings, T² and SPE with control limits, and VIP, all interactive:
- Bring in data your way - upload a CSV or Excel file, paste a public URL, or start from one of the bundled example datasets (LDPE, food consumption, solvents, NIR tablet spectra, and more).
- Understand each variable - click any column to see its histogram, run-order sequence, and summary statistics, and preview a transform (log, logit, ...) before you commit to it. Toggle the whole grid between raw and autoscaled values.
- Clean and shape the workset - set the primary identifier (auto-detected, with a synthetic row id when needed), mark X / Y variables, and exclude outlying samples or unwanted variables - all without altering your original data.
- Catch problems early - duplicate identifiers, non-numeric values in numeric columns, and missing data are flagged as you go.
- Fit PCA and PLS - choose the number of components and fit. For one-component models the plots collapse to a single, readable axis automatically.
- Interrogate the model - plot any pair of components (or sequence order), hover a score to read its SPE and T², and double-click or long-press any point to open its contribution plot and see which variables drive it. Long-press a loadings point to jump straight to that variable's raw data.
- Select and link across every plot - switch a plot into arrow or lasso mode and pick out points; the selection lights up the same rows (or variables) in every other plot - scores, loadings, T², SPE, VIP, and contributions - so a cluster you spot in one view is instantly visible everywhere. The lasso is forgiving on touch: a partial ring closes itself.
- Colour and size by what matters - shade or scale the score points by SPE or Hotelling's T² to make outliers jump out.
- Act on a selection - fork a new model variant straight from a brushed set: exclude the selected samples (or keep only them) and refit in one click, with the lineage recorded automatically.
- Keep a history - every model variant lands in the Hangar, and each one carries a Logbook recording its preprocessing, exclusions, and lineage, so you can branch a new variant from an existing one and compare. The running version is always shown, faint, in the top-right corner so you know exactly what you are looking at.
Your datasets and models are saved, so they survive a restart and are there when you come back.
You need Python 3.12+ and uv. Then:
uv run scorepilotThat boots the app and opens your browser at http://127.0.0.1:8000. Load a sample
dataset from the home page and start exploring.
Handy flags:
uv run scorepilot --host 0.0.0.0 --port 8080 --no-browserThe packaged app is a single process with no Node required at runtime: the Python server hosts both the API and the web UI.
ScorePilot has an optional in-app assistant, T²-D2, for help interpreting a model. It is off by default, bring-your-own-key, and the analysis never depends on it - the tool is fully usable without it.
ScorePilot is a FastAPI + SQLAlchemy backend with a SvelteKit (Svelte 5) + ECharts
frontend, wrapping the process-improve
chemometrics library for the numerics. The numerical core/ is pure and has no web or
database imports.
The frontend ships two standalone, domain-agnostic libraries that know nothing about
chemometrics and can be reused elsewhere: frontend/src/lib/grid/ (a dependency-free
data grid) and frontend/src/lib/plots/ (a linked-plots collection with
brushing-and-linking; see frontend/src/lib/plots/README.md).
- Architecture, the dataset / preprocessing-spec model, and the reusable data grid:
docs/ARCHITECTURE.md - A from-scratch local setup walkthrough:
docs/DEVELOPMENT.md - Deployment and publishing:
docs/DEPLOYMENT.md,docs/PUBLISHING.md
Common tasks:
uv sync # set up the environment
uv run scorepilot # run the app
uv run pytest # tests
uv run ruff check . # lint
uv run pyright # type check
cd frontend && npm install && npm run dev # frontend dev server (proxies /api)MIT. See LICENSE.