Before modelling: how to determine whether a sensor dataset is usable
A quality audit across acquisition, time, missingness, artefacts, protocol and retained coverage. Includes a synthetic session with contact loss, drift and event misalignment.
AmperieLabs
Technical notes built around failure cases.
Each note contains a concrete method, diagram, synthetic experiment, code pattern or review checklist. Publication frequency follows the material. There is no weekly commentary quota.
From Jupyter notebook to reproducible scientific pipeline
The point where a notebook needs schemas, configuration, tests, a clean run and a manifest. Includes a practical repository structure.
Why random windows inflate performance in physiological and sensor ML
Participant identity, overlapping windows and preprocessing leakage. Includes a grouped-validation experiment.
How to validate cell segmentation when ground truth is limited
Stratified annotation, object-level errors, merges, splits and metric choice for Cellpose, StarDist and classical baselines.
Sensor drift, event misalignment and motion artefact are different problems
Three failures that can produce similar-looking traces and require different evidence.
When machine learning is not the next step for a scientific dataset
Six situations where labels, protocol, sample structure or operating cost should stop the model build.