// the find
h1st-ai/h1st
Power Tools for AI Engineers With Deadlines
H1st is a Python framework from Arimo-Panasonic for industrial ML work where domain rules written by experts sit alongside trained models. Its core pieces are rule-based models, ML models, an Oracle that combines the two, and H1Flow, a graph for wiring steps together. It is aimed at data scientists on predictive maintenance, IoT and similar problems where there is too little history to train on.
The Oracle is the most useful idea here: encoded domain rules and a trained model live in one object, and the Azure IoT predictive maintenance example shows the rules covering cases the model has no data for. The trust code is split into describable, explainable, auditable and debiasable modules with LIME and SHAP explainers, so explanation is a separate concern rather than a notebook afterthought. Model storage sits behind a small interface with local and S3 backends, which is a sensible seam for moving artifacts without touching model code. The tests target the parts that break easily: the Oracle and time-series Oracle, fuzzy logic models, the model repository, and the H1Flow graph.
The last push was 31 July 2023, so this has been untouched for over three years. The Python 3.8 floor is end-of-life, and I would expect pinned dependencies to fight a current environment. On Windows it needs 64-bit Python plus Visual Studio Build Tools, which the README mentions in one line, so expect a rough first hour on a fresh machine. The repo README is mostly philosophy with the install steps after it, the real documentation lives on readthedocs, and most worked examples are notebooks, which are awkward to diff or run in CI.