NoCodeML V3: guest-first temporary AutoML workspace

[Rishikeshsanin]

Rebuilds NoCodeML V3 as a temporary, account-free AutoML utility while preserving the previous versions on dedicated recovery branches.

Product flow

Open → upload → explore → configure → train → compare → predict → export → leave.

No signup is required. Visitor datasets, generated models, training state, predictions and exports are not written to the application database.

Guest/privacy architecture

• cryptographically random anonymous session token
• raw token never used as a filesystem directory name
• isolated temporary workspace per session
• explicit Clear & Restart deletion
• browser close cleanup signal with refresh-safe grace period
• inactivity TTL cleanup
• active-training cleanup leases
• session isolation/path traversal tests
• public persistent auth/dataset/experiment/training/prediction routes removed
• PostgreSQL/Supabase not required by the public runtime
• Redis/Celery not required by the public runtime

ML workflow

• CSV / Excel / Parquet uploads with 100 MB bound
• EDA, missingness, statistics, correlations and visualizations
• ML Readiness guidance
• target/task/feature/model suggestions
• classification + regression
• Logistic/Linear Regression, Random Forest, XGBoost, LightGBM
• fitted sklearn preprocessing persisted with each temporary model
• safe unseen-category inference
• bounded in-process training pool
• progress + model-level results + best-model selection
• single + batch prediction

Downloads

• chart PNG with descriptive filenames
• EDA summary JSON
• statistics CSV
• missing-values CSV
• correlations CSV
• model comparison CSV
• training summary JSON
• feature importance CSV
• best model .joblib
• prediction CSV
• complete session ZIP

AI assistant

The Data Science Assistant is temporary-session scoped. Provider calls remain server-side and its automatic context is built from derived workspace metadata/results rather than raw uploaded rows.

Engineering / deployment

• React/TypeScript guest-only routing and header
• responsive guided workspace
• FastAPI guest-only public API
• Docker single-service ephemeral runtime
• non-root container
• /health and database-free /ready
• GitHub Actions frontend, backend, deployment-config, container and ARM64 ML validation
• Oracle Always Free ARM deployment stack with Caddy HTTPS and tmpfs visitor workspaces
• repository-root Vercel build configuration
• V2 preserved on legacy/v2-2026-08-22
• pre-guest V3 preserved on checkpoint/v3-rc-persistent-2026-08-22

Release gates

• final CI green on release head — run #270
• version promoted to 3.0.0
• production deployment configuration validated in CI
• public Oracle backend deployment
• Vercel frontend deployment
• live classification E2E
• live regression E2E
• mobile + desktop live QA
• v3.0.0 GitHub release/tag

The code release is ready for main; the remaining gates are provider-side deployment and live-environment verification.