Spline Preprocessing + LSTM/GRU Neural Forecasting.
Short-term noise and long-term dependencies — handled together.
A production-ready time-series forecasting pipeline that combines spline-based preprocessing with LSTM and GRU neural networks. Multi-input modeling, rolling-window cross-validation, CLI runner, FastAPI backend, React UI, and first-party agent ecosystem integration.
The project was structured as a phased build — each phase validated before proceeding, with comprehensive documentation, test suites, and operational gates at every stage.
Spline interpolation, scaling, and windowing pipeline with reproducible artifact persistence.
LSTM and GRU architectures supporting past observations, future-known variables, and static covariates.
Unified CLI training runner with rolling-window cross-validation and configurable hyperparameters.
Operational gates with smoke testing, run-verification, and circuit-breaker patterns for inference safety.
FastAPI backend exposing forecast endpoints and React UI for model monitoring, input management, and output visualization.
Backend compatibility with Tollama agent runtime, LangChain, and OpenClaw skills for integration into multi-agent forecasting workflows.
Spline interpolation for gap filling and noise smoothing before neural modeling — capturing the underlying trend without overfitting short-term variance.
Both LSTM and GRU architectures are supported, configurable via YAML. Each captures long-range temporal dependencies while handling varying sequence lengths.
Supports past observations, future-known variables (e.g., promotions, calendar features), and static covariates (e.g., store type) as separate model inputs.
Train, validate, and export models from a single CLI command with full artifact persistence — checkpoints, metrics, and config saved automatically.
Time-series cross-validation using a rolling-window methodology — no data leakage, realistic evaluation of model performance across historical regimes.
Smoke testing and run-verification gates ensure models meet quality thresholds before being promoted to production inference endpoints.
REST API for forecast endpoints with a React frontend for monitoring, input management, and visualizing predictions — deployable standalone or embedded.
First-party backend compatibility with the Tollama agent runtime, LangChain, and OpenClaw skills — plugs directly into multi-agent forecasting workflows.
TensorFlow-backed LSTM/GRU with spline preprocessing, served via FastAPI and monitored through a React UI. Structured for clean agent integration.
spline-lstm is designed to operate as a forecasting backend in multi-agent systems — compatible with the Tollama daemon, LangChain chains, and OpenClaw skill wrappers.
Register spline-lstm as a model family in the Tollama daemon — accessible via the same REST API and SDK used for TSFM models like Chronos and TimesFM.
Expose the FastAPI forecast endpoint as a LangChain tool — letting LLM agents call spline-lstm forecasts as part of reasoning chains.
Compatible with OpenClaw skill contracts — wrap as a structured skill with health, predict, and explain endpoints for Claude Code agent use.
Six phases complete. Agent-ready for integration into any forecasting workflow.