prime-platform

byKASAK PAREKH

Use this as your **complete final master prompt**: --- # FINAL MASTER PROMPT — Production Scientific AI/ML Prediction SaaS Platform Build a **production-ready, enterprise-grade, research-grade AI-powered scientific machine learning web platform** for: * Scientific prediction * Frequency extrapolation * Multi-model benchmarking * Scientific graph validation * Experiment tracking * Research export * Commercial SaaS deployment This platform must support: * THz datasets * S-parameter datasets * RF / Microwave datasets * Electromagnetic frequency-response datasets * Signal-processing datasets * General numerical scientific datasets The platform must be suitable for: * Universities * Research labs * Engineering companies * Industrial R&D * IEEE publication workflows * Commercial SaaS deployment --- # 1. Full Stack Architecture Frontend: * React + TypeScript * Responsive UI * Scientific dashboards * Real-time updates Backend: * Python + FastAPI Database: * PostgreSQL Caching: * Redis Background tasks: * Celery --- # 2. Authentication & Security Implement: * Register/Login * JWT authentication * Role-based access * Project isolation * Session security * Password hashing * Rate limiting * Audit logging * File validation * Input sanitization --- # 3. Supported ML Models Implement: * KNN Regressor * Decision Tree Regressor * Random Forest Regressor * AdaBoost Regressor * Gradient Boosting Regressor * XGBoost Regressor * Extra Trees Regressor * SVR * MLP Regressor Every model must preserve its own learning behavior. Never force identical curves across models. --- # 4. Dataset Upload & Validation Support: * CSV * Excel * TXT Automatically detect: * Input columns * Target columns * Frequency columns * Scientific measurement columns Validate: * Remove null rows * Remove corrupted rows * Remove duplicates * Validate numeric-only columns * Detect invalid scientific values Display: * Total rows * Valid rows * Removed rows * Feature columns * Target columns Reject invalid datasets. --- # 5. Training Pipeline Use proper ML workflow: Train/Test split: * 80/20 or configurable * Fixed random seed Never evaluate on training data. Automatically retrain when user changes: * Dataset * Features * Targets * Frequency range * Models * Preprocessing Never reuse stale models. --- # 6. Preprocessing Tree models: * Direct training Distance/Neural models: * KNN * SVR * MLP Must use: * StandardScaler or * MinMaxScaler Save preprocessing pipeline. --- # 7. Hyperparameter Optimization Use: * GridSearchCV or * RandomizedSearchCV Save best hyperparameters. Never rely only on defaults. --- # 8. Cross Validation Use: * 5-Fold or configurable K-Fold validation Display: * Mean score * Standard deviation * Stability score --- # 9. Evaluation Metrics For every model calculate: * MAE * MSE * RMSE * R² * MAPE Auto-rank models. Highlight best model. --- # 10. Overfitting Detection Compare: Training vs Testing performance. If large gap: Show: * Overfitting warning * Underfitting warning --- # 11. Prediction Pipeline (Mandatory Protocol) Every prediction must follow: User Input → Input Validation → Feature Processing → Scaling (if required) → Model Inference → Output Validation → Scientific Validation → Graph Rendering Never skip stages. --- # 12. User Input Compliance Whenever user changes: * Number of extrapolation points * Start frequency * End frequency * Frequency step * Manual frequency values * Target variable * Model selection Backend must: * Validate input * Accept valid scientific values * Recalculate predictions from scratch Never ignore user values. Never reuse cached outputs. --- # 13. Scientific Extrapolation Rules When user increases points: For every model: Generate new X values. For every new point: Run actual `.predict()` inference. Never: * Copy old values * Stretch old curves * Shift graph lines * Duplicate predictions * Use fake frontend interpolation Every point must come from real ML inference. --- # 14. Scientific Curve Rules For THz / S-parameter / frequency-domain data: Preserve: * Frequency continuity * Slope continuity * Peaks * Valleys * Resonance dips * Curvature transitions Predictions must never: * Drop vertically * Rise unnaturally * Become flat randomly * Repeat identical values * Zig-zag unnaturally * Show non-physical spikes * Show discontinuous jumps Reject unrealistic curves automatically. Never display invalid scientific predictions. --- # 15. Boundary & Confidence Protection If user enters values outside training range: Detect out-of-distribution extrapolation. System must: * Warn user * Mark low-confidence regions * Highlight extrapolated zones Never present low-confidence predictions as guaranteed truth. --- # 16. Model-Specific Prediction Rules KNN Regressor: * Follow nearest-neighbor logic SVR: * Respect kernel boundaries MLP Regressor: * Preserve nonlinear learned behavior Tree models: * Respect learned split boundaries Never force one model’s prediction behavior onto another. --- # 17. Graph System Build: 1. Actual vs Predicted 2. Frequency Extrapolation 3. Residual/Error Plot 4. Model Comparison Plot Use: * Plotly --- # 18. Graph Rendering Rules Actual: * Solid line * Thick stroke * Circle markers Predicted: * Dashed line * Thin stroke * Square/Triangle markers If overlap occurs: Apply visualization-only separation: * z-order * Marker differentiation * Adaptive spacing Never modify real values. Legend: Auto-place in emptiest region. Never cover curves. --- # 19. Dynamic Graph Updates Whenever user changes points: Regenerate all graphs using fresh predictions. Never reuse cached Y-values. --- # 20. Prediction Validation Layer Before graph rendering validate: * No NaN * No Infinity * No repeated fake sequences * No unrealistic spikes * No vertical collapse * No invalid derivative discontinuity If anomaly detected: Block rendering. Re-run: * Preprocessing * Inference * Validation --- # 21. Explainability Provide: Feature importance. Support: * SHAP --- # 22. Auto Model Recommendation Recommend best model based on: * Accuracy * Stability * Extrapolation quality * Scientific smoothness Not only R². --- # 23. Model Persistence Save models using: * Joblib Save: * Model * Scaler * Hyperparameters * Metrics * Metadata * Timestamp --- # 24. Experiment Tracking Save: * Dataset version * Model version * Random seed * Hyperparameters * User settings * Timestamp Support reproducibility. --- # 25. Export System Support: * CSV * Excel * JSON * PDF * PNG * SVG * TIFF Include: * Model info * Metrics * Graphs * Dataset info * Timestamp Publication quality exports only. --- # 26. Large Dataset Handling For 10k+ rows: * Chunk processing * Memory optimization * Background execution * Safe caching Prevent crashes. --- # 27. Fault Tolerance If one model fails: * Log error * Skip failed model * Continue remaining models Never crash full platform. --- # 28. Live Logs Show: Dataset uploaded... Validation complete... Preprocessing... Training... Hyperparameter tuning... Prediction... Graph rendering... Export ready... --- # 29. Notifications Notify users when: * Training completed * Export completed * Model failed * Dataset invalid Support: * Email * In-app notifications --- # 30. Admin Dashboard Build controls for: * Users * Storage * API usage * Failed jobs * Billing * System logs --- # 31. API Documentation Auto-generate: * Swagger UI --- # 32. SaaS Billing Architecture Prepare future billing with: * Stripe Plans: * Free * Research * Pro * Enterprise --- # 33. Deployment Support deployment on: * Docker * Render * Railway * Microsoft Azure * Amazon Web Services Must support: * Environment variables * Secure secrets * Monitoring * Logging * Production scaling --- Build this as a **real scientific AI platform**, not a demo or fake graph generator. Every prediction must follow actual machine learning rules, scientific continuity, model-specific behavior, and user-entered values.