¶

Neural codecs for wearable and physiological signals on edge devices.

compressionKIT helps teams compress continuous sensor waveforms before they become a memory, radio, battery, or cloud-ingestion problem. It is designed for wearable and edge signals such as PPG, ECG, IMU, and future time-series modalities where preserving waveform utility matters.

The package provides reusable codec blocks, deploy packaging, runtime loading, and validation scorecards. The v1 release artifacts focus on PPG and ECG, but the architecture is intentionally broader than any single method or modality.

What compressionKIT is¶

compressionKIT is a toolkit for building, evaluating, packaging, and deploying compression codecs for wearable signals.

It includes:

• neural codec models and quantization components
• DSP and hybrid codec stages where they are useful
• preprocessing and augmentation blocks
• physiological and signal-fidelity scorecards
• deploy package generation with manifests, checksums, reference vectors, and C headers
• runtime loaders for local and HuggingFace artifacts
• golden release automation for supported PPG and ECG operating points

It is not a general physiological inference framework. It does not try to own arrhythmia classification, sleep staging, activity recognition, or downstream clinical decisions. It focuses on the codec layer: reducing data movement while preserving the signal content downstream systems need.

Where compressionKIT fits¶

compression sits between the sensor and everything downstream: local storage, radio transport, gateway buffering, cloud ingestion, analytics, and model training. That position makes the codec a product boundary, not only a model choice. A release package needs to be small enough for edge deployment, explicit enough for validation, and stable enough that future codec versions can be compared against previously supported ones.

What current goldens show¶

The first release-grade packages exercise the artifact contract on PPG and ECG, but the top-level question is not a single PRD curve. A customer needs to choose an operating point from compression level, physiology preservation, robustness, stitching behavior, and deploy footprint together.

The generated customer evidence summary is the compact entry point for this decision. Detailed plots now belong on the modality and CR-vs-fidelity pages, where they can be read with the matching sample counts, noise buckets, and reproduction links.

How to read the evidence¶

The headline numbers are a starting point, not the whole story. compressionKIT reports each supported codec through the same evidence surfaces so users can see what changed, what stayed comparable, and which signal conditions were tested.

The guiding rule is simple: comparisons belong in generated scorecards, noise/artifact sweeps, and CR-vs-fidelity pages, not in one-off prose. When a new method or release arrives, the tables are regenerated from the supported artifacts so old and new versions remain comparable.

Guiding principles¶

These principles keep the toolkit practical without making every experiment use the same control flow.

Why it matters¶

High-level getting started¶

Start with the lightest path that answers your question.

Minimal runtime example:

import numpy as np
from compressionkit.runtime import load_codec

codec = load_codec("Ambiq/compressionkit-ppg-4x-v1.0")
frame = np.zeros(codec.frame_size, dtype=np.float32)

encoded = codec.compress(frame)
reconstructed = codec.decompress(encoded)

Install locally:

uv pip install "compressionkit[hf]"

# Or, from a source checkout:
uv sync --extra hf

Current v1 artifacts¶

The first release-grade packages exercise the common artifact contract across PPG and ECG:

For method details, see Methods. For the philosophy behind blocks, ready-made experiments, and goldens, see Experiment Architecture.

Where to go next¶