@paragondao/bagle-sdk

SDK for the BAGLE (Brain-AI General Learning Encoder) API.

Encode any biosignal — breathing, EEG, accelerometer, heart rate, typing cadence, gait — into 128 universal coefficients (512 bytes). The encoder is signal-agnostic: it only sees numbers.

Install

npm install @paragondao/bagle-sdk

Quick Start

import { BagleClient } from '@paragondao/bagle-sdk';

const client = new BagleClient();

// Encode any time-series into 128 coefficients
const result = await client.encode([1.2, 3.4, 5.6, 7.8, /* ... */]);
console.log(result.encoding); // 128 numbers
console.log(result.dim);      // 128
console.log(result.latency_ms); // ~3ms

// Compare two encodings
const sim = await client.similarity(encodingA, encodingB);
console.log(sim.similarity); // 0.0 to 1.0

Client-Side Encoding (Offline)

For encoding without hitting the API:

import { gleEncodeSignal, cosineSimilarity } from '@paragondao/bagle-sdk';

// Any numeric array — accelerometer, heart rate, breathing, typing intervals
const coefficients = gleEncodeSignal([1.2, 3.4, 5.6, 7.8, /* ... */]);
console.log(coefficients.length); // 128

// Compare locally
const score = cosineSimilarity(coefficientsA, coefficientsB);

API Reference

BagleClient

const client = new BagleClient({
  baseUrl: 'https://bagle-api.fly.dev', // default
  apiKey: 'your-key',                    // optional
  timeout: 10000,                        // ms, default
});

client.encode(features, modality?)

Encode a numeric array into 128 GLE coefficients.

client.similarity(encodingA, encodingB)

Cosine similarity between two 128-dim encodings. Returns { similarity, distance }.

client.health()

Check API status and model availability.

Client-Side Functions

gleEncode(audioSamples, sampleRate?)

Encode PCM audio (16kHz default) using 1-second windows with identity band extraction (20-1000Hz).

gleEncodeSignal(samples, opts?)

General-purpose encoder for any numeric time-series. Options: windowSize, bandMin, bandMax.

cosineSimilarity(a, b)

Compute cosine similarity between two coefficient vectors.

What Can You Encode?

Every phone sensor produces a time-series. Every time-series can be GLE-encoded:

| Signal | Source | Example App | |--------|--------|------------| | Breathing audio | Phone mic | Crisis screening, stress detection | | Heart rate | Camera PPG | Fitness, sleep quality | | Accelerometer | Motion sensor | Gait analysis, fall detection | | Gyroscope | Rotation sensor | Tremor detection, balance | | Typing cadence | Keyboard | Cognitive load, fatigue | | EEG | MUSE headband | Consciousness classification | | Temperature | Sensor | Fever screening | | Barometric pressure | Barometer | Altitude sickness |

The network doesn't care what signal produced the 128 numbers.

License

MIT