Ridge

Evidence-backed analysis for spreadsheets.

Ridge turns spreadsheets and structured files into traceable findings, quality diagnostics, statistical evidence, and clear explanations.

Live: ridge-data.vercel.app — drop a file in and get results. No account, no API key.

Most "AI for your spreadsheet" tools hand a language model a few rows and ask it to be insightful. The model then invents statistics it cannot see, reports correlations without saying how many observations they rest on, and treats blank cells as zeroes — which quietly drags means toward zero and manufactures relationships that are not in the data.

Ridge inverts the order. Everything numeric is computed deterministically first — statistics, data quality, correlations, evidence — and that is the product. A language model is offered afterwards, optionally, and its only job is to explain evidence that already exists.

/about redirects to /. A share link of the form /?a=<id> forwards to /app?a=<id>, so links created before the landing/workspace split still open.

The workspace at /app after Explore the sample: deterministic evidence, statistics and quality diagnostics, with AI interpretation still un-run.

Evidence and statistics	Data quality

Up to 10 files, 4 MB total per request (a Vercel body limit, enforced on both the client and the server). Larger files — to 25 MB — go through the URL box, which fetches server-side.

Everything below is computed on the server. No model is involved and nothing is sent to Anthropic.

• Structural reading — where the header actually starts, and which rows restate other rows instead of being observations. A title block above the header and a TOTAL row at the bottom are excluded before anything is computed, and every excluded row is reported with the reason. When the file does not settle the question, Ridge says so rather than guessing, and you can correct the header row or put a row back. A sheet whose shape reads as transposed — field names down the first column, records across the columns — is declared uncertain instead of being averaged in silence.
• Column statistics — valid/missing/invalid counts, coverage, min, max, mean with a 95% t interval, median, standard deviation, quantiles, and IQR outlier fences.
• Categorical profiles — top values ranked by frequency with counts and percentages, unique counts, and whether a field reads as an identifier, a category, or high-cardinality text.
• Date profiles — valid/invalid counts, range, bucketed trend, period-over-period change, gaps, and irregular intervals.
• Correlations — Pearson and Spearman over pairwise-complete observations, each reported with method, n, coverage, strength and caveats — and each drawn: the paired observations travel with the coefficient as a scatter (or a density grid for large pairings), with an at-a-glance matrix over every numeric column pair.
• Charts for every column — a histogram, frequency or trend chart for each column with a computed full-file aggregate, numeric spread strips (five-number summaries with IQR fences), per-column completeness tracks, and one-click PNG export of any chart.
• Inference and change detection — Welch mean intervals, chi-square with Cramér's V, two-sample KS distribution shifts, and exploratory robust level shifts over dated targets.
• Data quality — a weighted health grade with per-column issues, missingness, type consistency and duplicate detection.
• Evidence Mode — see below.
• Exports — JSON, a printable HTML report, and any chart as a PNG.
• Comparison mode — baseline/current schema drift, quality movement, descriptive deltas, mean intervals, and distribution shifts for two files.

Add your own Anthropic key (⚙ in the top bar) to unlock:

• Explain with Claude — an interpretation of the evidence already computed.
• Follow-up questions — ask about the dataset in plain language.
• Cross-file synthesis — themes and differences across a multi-file upload.
• Document reading — PDFs, Word and PowerPoint get real analysis rather than an extracted-text preview.

Models: Claude Sonnet 5 (default, balanced), Claude Opus 5 (most capable), Claude Haiku 4.5 (fastest). The key is used per request and never stored server-side — see PRIVACY.md.

Every material finding is a structured object rather than a sentence:

{
  "claim": "revenue and spend rise together (spearman +0.9068)",
  "metric": "spearman_rho",
  "value": 0.9068,
  "columns": ["revenue", "spend"],
  "method": "spearman correlation, pairwise-complete observations",
  "sampleSize": 90,
  "coverage": 98.9,
  "strength": "very strong",
  "caveat": "Pearson and Spearman disagree, suggesting a non-linear or outlier-driven pattern",
  "statistics": null,
  "engineVersion": "1.3.0",
  "provenance": {
    "formula": "ρ = Pearson correlation of the paired value ranks",
    "includedRows": 90,
    "excludedRows": 1,
    "sourceRows": [{ "rowNumber": 1, "values": { "revenue": 100, "spend": 45 } }]
  }
}

Pick an optional target column and the engine focuses on that outcome: group comparisons across categories with a standardized effect size and Welch interval, categorical associations with Cramér's V, correlations involving the target, missingness impact, and trends over time.

The AI layer may summarize and contextualize these objects and must carry their caveats forward. It cannot add numbers they do not contain. Deterministic results and AI interpretation are labelled distinctly everywhere they appear, including in the exported report.

• Files are processed in memory and nothing is retained after the response.
• Nothing is stored server-side unless you opt in per analysis with "Save this analysis to history and enable a share link", which is off by default. A configured database is capability, not consent.
• Saved analyses can be deleted from the history list.
• Your API key lives in session storage by default (cleared with the tab), or local storage only if you tick Remember this key on this device. It is sent with each AI request to this backend, which forwards it to Anthropic. It is never stored server-side or written to logs.

Full detail: PRIVACY.md and the live /privacy page.

Use docs/PILOT_PLAYBOOK.md to recruit and observe five people with recurring finance, operations, or analytics spreadsheet work. It defines the session protocol, privacy boundary, return-use metric, and decision thresholds. Non-confidential product feedback can be submitted through the repository's Pilot feedback issue form.

git clone https://github.com/MohammedAlkindi/Ridge && cd Ridge
npm ci
npm run dev            # http://localhost:3000

Click Try sample data — no configuration needed.

All optional; the app runs with none of them set.

Copy .env.example to .env to set them locally.

public/            static app (/, /app, /privacy, /docs) + sample data
src/routes/        analyze · ask · explain · history · health
src/services/      anthropic · remoteFile · history          (all I/O)
src/analytics/     values · stats · dates · correlations ·
                   evidence · sample · profile               (pure)
src/parsers/       spreadsheet · document · structured
api/index.js       Vercel serverless entry

Routes validate and orchestrate; services own every external call; analytics and parsers are pure and unit-tested. Errors become AppError and reach the client only as { error, code, requestId }. Details in docs/ARCHITECTURE.md; endpoint reference in docs/API.md.

npm test              # unit + API integration (vitest) — no network, no key
npm run test:browser  # Playwright journeys, desktop + mobile
npm run test:all      # both

The suite mocks the Anthropic SDK; it never needs a real key and never makes a provider call.

Deployed on Vercel as static public/ plus one serverless function (api/index.js). vercel.json disables framework detection, enables cleanUrls, redirects /about → /, and forwards /?a=<id> to /app so share links from the previous layout keep working.

vercel --prod

Set ANTHROPIC_API_KEY, SUPABASE_URL and SUPABASE_KEY in the Vercel project only if you want a server fallback key and history.

Ridge also ships a non-root, health-checked OCI image for private deployments:

docker build -t ridge:local .
docker run --rm -p 3000:3000 ridge:local

Data-flow modes, scaling limits and the production checklist are documented in docs/DEPLOYMENT.md.

Semantic versioning. Every release has a matching package.json version, a CHANGELOG.md entry, an annotated vX.Y.Z tag, a GitHub Release, green checks and a verified production deployment. Process: docs/RELEASING.md.

See CONTRIBUTING.md. Security policy: SECURITY.md. Roadmap: ROADMAP.md.

MIT — see LICENSE.