SheetPredict

SheetPredict turns a spreadsheet into a validated, explained, deployable model in minutes — and shows its work.

It is an AI-guided platform for analysts who are not ML engineers: upload a CSV or Excel file, and a guided workflow takes it through profiling, preparation, feature selection, AutoML training, evaluation, prediction and deployment, without writing code.

What it does

Eight guided stages (apps/frontend/lib/types/workflow.ts):

1. Data loading. CSV/XLSX upload (100 MB cap, chunked for large files), PII scanning on the in-app upload, and sample datasets for onboarding.
2. Data profiling. Column statistics, charts, data-quality issues with suggested fixes, and an AI summary.
3. Data preparation. A visual pipeline of the four transformations the engine runs (remove duplicates, trim whitespace, drop missing, fill missing), with preview, undo/redo, version history and exportable Python code.
4. Feature engineering. Choose features, with identifiers, dates and empty columns detected and excluded automatically.
5. Model training. AutoML over several algorithms with optional tuning, bounded per plan tier.
6. Model evaluation. Leaderboard against a baseline, metrics, SHAP explanations and a model report that states only what was recorded.
7. Prediction. Single-record and batch predictions that warn on unseen or missing inputs.
8. Deployment. A REST endpoint secured by API keys and rate limits, drift monitoring, a generated SDK, and self-contained Python/Docker exports.

Every AI feature has a rule-based fallback, so the app works without an OpenAI key. Billing (Stripe) is optional: without keys every account runs on the free tier.

Not supported yet: database or cloud-storage connectors, multi-file joins, uploads above 100 MB.

Repository layout

apps/
  backend/    FastAPI API: datasets, transformations, training, serving, billing
  frontend/   Next.js app (App Router, NextAuth v5, Tailwind v4)
  mcp/        MCP server exposing the EDA summary tool to the backend
docs/
  adr/        Architecture decision records (billing, plan limits and pricing)
  dev/        Local development, backend test infrastructure, E2E testing
  ops/        Staging deployment, credential rotation, erasure/backup, artifact integrity
  product/    Positioning and public-surface decisions
scripts/      Local setup, deploy, CI, quality-gate and ops scripts

Tech stack

• Frontend: Next.js 16, React 19, TypeScript, Tailwind CSS 4, NextAuth v5 (Google and GitHub sign-in), Jest, Playwright
• Backend: Python 3.13, FastAPI, Pydantic, Beanie (MongoDB ODM), pandas, scikit-learn, XGBoost, LightGBM, SHAP, Optuna
• Data: MongoDB, Redis (optional: cache and shared rate limiting; without it the limiter is process-local), S3 (LocalStack locally)
• AI: OpenAI, behind a circuit breaker, per-tenant quotas and a global daily ceiling
• Tooling: uv, npm, Docker, GitHub Actions

Getting started

Prerequisites: Docker, Python 3.13 with uv, and Node ^22.13.0 || >=24.

# MongoDB on :27017 and LocalStack S3 on :4566, with the dev bucket created
./scripts/local-setup.sh

# Backend on :8000
cd apps/backend
cp .env.example .env          # then apply the Mongo overrides in docs/dev/LOCAL_DEVELOPMENT.md
uv sync
uv run uvicorn app.main:app --reload --host 127.0.0.1 --port 8000

# Frontend on :3000 (in a second shell)
cd apps/frontend
cp .env.local.example .env.local   # NEXTAUTH_SECRET must equal the backend's
npm install
npm run dev

apps/backend/.env.example sets SKIP_AUTH=true, which maps every request to a single dev user. The backend refuses to start with it unless ENVIRONMENT is development or test. docs/dev/LOCAL_DEVELOPMENT.md has the full setup, including real sign-in locally.

Testing

# Backend: the PR gate (needs MongoDB on :27017)
cd apps/backend && PYTHONPATH=. uv run pytest tests/ -m "not integration and not performance"

# Frontend
cd apps/frontend && npm test && npm run type-check && npm run lint

# E2E against the real stack
cd apps/frontend && npm run test:e2e:smoke

# MCP server
cd apps/mcp && uv run pytest tests/

.github/workflows/ci.yml runs all of these on every pull request; CI Success is the required check. See docs/dev/TEST_INFRASTRUCTURE.md and docs/dev/E2E_TESTING_GUIDE.md.

Deployment

main deploys to staging with .github/workflows/deploy.yml, using docker-compose.staging.yml behind nginx (nginx-staging.conf). The environment templates are .env.staging.example and .env.production.example at the repo root. Runbooks are in docs/ops/, starting with STAGING_DEPLOYMENT_GUIDE.md.

Contributing

Conventions, architecture rules and the guard tests that enforce them are in CLAUDE.md. Work on a feature branch and open a pull request against main.

License

Licensed under the Apache License, Version 2.0. See NOTICE for attribution.

Copyright © 2025-2026 Noatak Enterprises, LLC, dba Bria Strategy Group

Author

Frank Bria · frankbria.com · @FrankBria18044