kairos
Learning when to execute blockchain transactions.
Deep learning for low-fee transaction timing.
// overview
KAIROS learns from finalized EVM block history to choose a low-base-fee block within a fixed future horizon. It builds causal samples from Blockweaver datasets, compares sequence models on Ethereum, Polygon, and Avalanche, and scores their timing decisions with per-origin economic metrics.
The project extends the temporal experiment from SPICE and underpins the IEEE CCNC 2027 paper on KAIROS. Staged feature, context, HPO, and horizon studies run as Slurm GPU campaigns, and the selected models ship in an Expo/ExecuTorch demo.
// how it works
The Python pipeline resolves UUID-addressed Blockweaver datasets, builds causal block features and historical windows, and trains LSTM, Transformer, and Transformer-LSTM models with PyTorch Lightning. Candidates are selected by validation optimality gap, and held-out analysis adds a fixed-deadline rolling policy and time-clustered bootstrap intervals.
Servatus handles durable campaigns, atomic publication, and OpenSSH/Slurm/Apptainer submission. An isolated exporter converts twelve selected models to XNNPACK ExecuTorch programs with parity checks, and the Expo app runs them on-device over public RPC without a server fallback.
// capabilities
Causal temporal samples
Uses strict closed-parent history and fixed block-count horizons to prevent future-data leakage.
Dual-head sequence models
Compares three model families that predict both a target block offset and its minimum base fee.
Held-out economic evaluation
Reports optimality gap and base-fee and P50 fee-inclusive savings with UTC-hour bootstrap intervals.
On-device inference
Bundles twelve parity-checked models, three chains by four horizons, into a serverless Expo app.