ROADMAP.md

TinyML Project Roadmap & Todo List

Phase 1: Core Architecture & Reliability (Completed/In Progress)

• Refactor Server State: Replace global variables with thread-safe ModelManager singleton.
• Thread Safety: Implement std::mutex locking for all model access and scenario updates.
• Lazy Initialization: Models and scenarios are created on-demand to save resources.
• Numeric Precision Unification: Migrate codebase to consistent precision (template-based or unified float32). Currently Network uses double while SIMD operations use float.
• Autograd System: Replace manual backpropagation logic with a proper Tensor class supporting automatic differentiation (similar to PyTorch).

Phase 2: Web-Deployable Demos (Portfolio Readiness)

Implemented

• Probabilistic Forecasting:
  • ForecastScenario generates P10/P90 quantile bands.
  • Visualization of uncertainty intervals in script.js.
• Attention Visualization:
  • AttentionScenario implements time-series attention mechanism.
  • Interactive heatmap and query-key connection visualization.

Planned Demos (To Implement)

• Physics-Informed Neural Networks (PINN):
  • Enhance HeatEquationScenario to toggle between "pure NN" and "physics-informed" modes.
  • Visualize stability differences in real-time.
• Interactive Bayesian Regression:
  • Upgrade BayesianScenario to allow dragging training points.
  • Visualize predictive variance reacting to data density in real-time.
• Spatiotemporal Graph Modeling:
  • Implement a grid/network map scenario (e.g., energy grid).
  • Animate demand shocks propagating through the graph.
• Conditional Generative Models:
  • Enhance GenerativeScenario to support conditional generation (e.g., demand curves conditioned on temperature).
• Reinforcement Learning (Energy Trading):
  • Create a battery arbitrage environment (charge/discharge/hold based on price).
  • Visualize policy evolution and cumulative reward.

Phase 3: Developer Experience & DevOps

• Documentation:
  • Architecture.md: High-level system design.
  • API.md: HTTP endpoint documentation (/run, /train, /scenario/*).
• Frontend Modernization:
  • Refactor script.js into ES6 modules (API layer, Charting layer, UI logic).
  • Consider moving to a lightweight framework (React/Vue) if complexity grows.
• Hyperparameter Tuning UI:
  • Add controls for Learning Rate, Batch Size, and Optimizer selection in the web interface.
• CI/CD & Deployment:
  • Add Dockerfile for containerized deployment.
  • Set up GitHub Actions for automated testing and build verification.

Phase 4: Advanced Features

• Rolling-Origin Evaluation: Add backtesting capability to the forecasting playground.
• Model Export/Import: Allow saving trained weights to JSON/binary and loading them back.
• WebAssembly (WASM) Port: Compile the C++ core to WASM to run inference entirely in the browser (client-side).