Trillboards Sensing SDK for Android

agent-core is the Android sensing primitive that powers Trillboards on-device audience intelligence: face detection, audio classification, speech recognition with diarization, BLE / WiFi / mDNS / SSDP / ARP / HTTP discovery, native sensor reads, and Vertex multimodal embeddings. Targets Android 8.0+ (minSdk=26), built against compileSdk=35.

Apache 2.0. Anonymous-readable via https://maven.trillboards.com/ — no tokens, no GitHub account. This repo is the Apache-2.0 source mirror.

Install

settings.gradle.kts:

dependencyResolutionManagement {
    repositories {
        google()
        mavenCentral()
        maven { url = uri("https://maven.trillboards.com/") }
    }
}

app/build.gradle.kts:

dependencies {
    implementation("com.trillboards:agent-core:1.1.+")            // floats on the latest 1.1.x

    // Optional — on-device speech (Moonshine ASR). Omit to ship without it.
    implementation("com.trillboards.bundled:sherpa-onnx:1.12.26")
}

Pin an exact version (e.g. 1.1.4) for reproducible builds.

Quick start

Foreground — SDK embedded in your on-screen player app. The camera follows app visibility via ProcessLifecycleOwner automatically; no service needed:

import com.trillboards.sdk.TrillboardsSensingSdk

TrillboardsSensingSdk.start(
    context = applicationContext,
    partnerApiKey = "YOUR_PARTNER_API_KEY",
)

Background / headless — sensing runs in its own app with no visible Activity (e.g. beneath a separate ad-player). Pass your FOREGROUND_SERVICE_TYPE_CAMERA LifecycleService as lifecycleOwner so the camera stays bound while backgrounded (start the service from a visible Activity per the Android 14 while-in-use rule):

import android.os.Build
import androidx.lifecycle.LifecycleService
import com.trillboards.sdk.SensingSdkConfig
import com.trillboards.sdk.TrillboardsSensingSdk

class CaptureService : LifecycleService() {
    override fun onCreate() {
        super.onCreate()
        val fgsType = FOREGROUND_SERVICE_TYPE_CAMERA or FOREGROUND_SERVICE_TYPE_MICROPHONE
        if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.Q) {   // typed startForeground is API 29+
            startForeground(NOTIF_ID, notification, fgsType)
        } else {
            startForeground(NOTIF_ID, notification)              // type comes from the manifest on API 26-28
        }
        TrillboardsSensingSdk.start(
            context = applicationContext,
            partnerApiKey = "YOUR_PARTNER_API_KEY",
            config = SensingSdkConfig(deviceId = "YOUR_DEVICE_ID"),
            lifecycleOwner = this,
        )
    }

    override fun onDestroy() {
        TrillboardsSensingSdk.stop(this)
        super.onDestroy()
    }
}

deviceId ties the install to a screen you registered via POST /v1/partner/device (the same value you used as external_id). When null, the SDK uses a hardware fingerprint and sensing runs unattributed until the device is bound.

What ships in the AAR

• Face: FaceXFormer demographics (age / gender), gaze estimation, per-track dwell + emotion
• Audio: MediaPipe AudioClassifier (yamnet), Moonshine ASR with speaker diarization, VAD
• Discovery: BLE scan (+ UWB / Auracast / ChannelSounding on API 36+), WiFi BSSID/SSID hashing, mDNS / SSDP / ARP / HTTP probes
• Native sensors: ambient light, barometer, IMU
• Cloud inference: Vertex multimodal embeddings via the Trillboards API (no GCP credentials on device)

Configuration

SensingSdkConfig exposes ~30 knobs (toggle each sensing primitive, sampling rates, diarization, FaceXFormer, debug logging). Common integrations:

• Full (camera + mic + face/audio) — defaults; the SDK adapts to device capabilities.
• Audio-only — SensingSdkConfig(faceDetectionEnabled = false).
• Discovery-only — SensingSdkConfig(faceDetectionEnabled = false, audioClassificationEnabled = false).

Privacy

• All face / audio / image inference happens on-device. No raw frames or audio buffers leave the device.
• Cloud calls send only structured features (counts, attention level, transcripts when diarization is enabled) — never images.
• GDPR / CCPA opt-out is per-screen via the Trillboards portal; the SDK respects opt-out flags from the heartbeat response.

License

Apache License, Version 2.0. See LICENSE. The bundled ML models (Moonshine ASR, MediaPipe AudioClassifier, pose landmarker, yamnet) ship under their respective licenses (Apache 2.0).

Reporting issues

File an issue on this repo, or email developers@trillboards.com.