Solutions

Generic Android Application Architecture

The Android application is designed with a clear separation of architectural responsibilities:

• Java/Kotlin → UI, Camera/Audio integration, Lifecycle management

• JNI → Minimal bridging layer between managed and native code

• C++ → Pre‑processing, Inference, and Post‑processing logic

• QAIRT SDK → Execution and hardware acceleration on CPU, GPU, or DSP using SNPE/QNN/GENIE APIs

Detailed Stage Breakdown

1. Java Layer (Application Logic)

Responsibilities:

• Camera capture (Android Camera2 API / CameraX)

• Frame acquisition (Bitmap / YUV)

• UI rendering (bounding boxes, labels)

• Threading & lifecycle

Typical data passed to JNI:

• Image buffer - Frame width / height, format

• User Prompt

• Hardware acceleration (CPU / GPU / DSP)

Receiving data from Native Layer:

• Post process detection outputs (bounding boxes, class labels, and confidence scores) and map results to screen coordinates

• Text/Image/Audio from GenAI

2. JNI Layer (Bridge Only)

Responsibilities:

• Minimal logic

• Data marshaling only

• No preprocessing or math

JNI exists to:

• Avoid SDK dependency in Java

• Keep performance-critical logic native

3. C++ Processing

Preprocessing:

Responsibilities:

• Resize to model input size

• Color conversion (YUV → RGB)

• Normalization / scaling

• Tensor layout conversion

This stage ensures:

• Input matches model expectations exactly

Inference:

Responsibilities:

• Load model once at startup

• Select runtime (CPU / GPU / DSP)

• Allocate buffers

• Execute inference per frame

This stage uses:

• SNPE APIs

• QNN APIs

• Genie APIs

Postprocessing:

Responsibilities:

• Decode raw output tensors

• Apply confidence thresholds

• Perform NMS (if needed)

• Convert model output → objects

Produces:

• Bounding boxes, Class IDs, Confidence scores for Object Detection Models

• Text for NLP

Why This Architecture Works Well

• Performance: Heavy work stays native

• Portability: Swap models without UI changes

• Scalability: Same flow for detection, segmentation, pose

QIDK Solutions

This document describes all end-to-end AI solutions provided in the Qualcomm Innovators Development Kit (QIDK). Solutions are grouped based on the AI stack used: SNPE, QNN, and GenAI (GENIE).

SNPE Solutions

NLPSolution1 – Question Answering

Description

End-to-end Question Answering pipeline demonstrating on-device NLP inference using Qualcomm AI accelerators.

Model Used

• ELECTRA – Question Answering

Model On-boarding

• ELECTRA Model On-boarding

Android App Source Code

• https://github.com/quic/qidk/tree/master/Solutions/NLPSolution1-QuestionAnswering/QuestionAnswering

NLPSolution2 – Sentiment Analysis

Description

Sentiment analysis example showcasing NLP inference optimized for Snapdragon platforms.

Model Used

• MobileBERT – Sentiment Classification

Model On-boarding

• MobileBERT Model On-boarding

Android App Source Code

• https://github.com/quic/qidk/tree/master/Solutions/NLPSolution2-SentimentAnalysis/SentimentAnalysis

NLPSolution3 – Automatic Speech Recognition (Whisper)

Description

Automatic Speech Recognition pipeline running fully on-device using Whisper and Qualcomm AI accelerators.

Model Used

• Whisper – Automatic Speech Recognition

Model On-boarding

• ASR Whisper Model On-boarding

Android App Source Code

• https://github.com/quic/qidk/tree/master/Solutions/NLPSolution3-AutomaticSpeechRecognition-Whisper/Android_App_Whisper

VisionSolution1 – Object Detection (YOLO NAS)

Description

High-performance object detection using YOLO NAS optimized for mobile AI acceleration.

Model Used

• YOLO NAS

Model On-boarding

• YOLONAS Model On-boarding

Android App Source Code

• https://github.com/quic/qidk/tree/master/Solutions/VisionSolution1-ObjectDetection-YoloNas/app

VisionSolution1 – Object Detection (DETR)

Description

Transformer-based object detection using DETR on Snapdragon platforms.

Model Used

• DETR – Detection Transformer

Model On-boarding

• DETR Model On-boarding

Android App Source Code

• https://github.com/quic/qidk/tree/master/Solutions/VisionSolution1-ObjectDetection-DETR/app

VisionSolution2 – Image Super Resolution

Description

Image super-resolution pipeline optimized for on-device inference.

Model Used

• SESR – Single Image Efficient Super Resolution

Model On-boarding

• SESR Model On-boarding Notebook

Android App Source Code

• https://github.com/quic/qidk/tree/master/Solutions/VisionSolution2-ImageSuperResolution/superresolution

VisionSolution3 – Image Enhancement

Description

Image enhancement solution focusing on low-light and visual quality improvement.

Model Used

• EnlightenGAN – Image Enhancement

Model On-boarding

• EnlightenGAN Model On-boarding

Android App Source Code

• https://github.com/quic/qidk/tree/master/Solutions/VisionSolution3-ImageEnhancement/enhancement

VisionSolution4 – Pose Estimation

Description

Human pose estimation pipeline optimized for Snapdragon devices.

Model Used

• HRNet – Human Pose Estimation

Model On-boarding

• HRNet Model On-boarding Notebook

Android App Source Code

• https://github.com/quic/qidk/tree/master/Solutions/VisionSolution4-PoseEstimation

QNN Solutions

VisionSolution1 – Object Detection (QNN based YOLO NAS)

Description

YOLONAS object detection implemented using Qualcomm AI Engine Direct (QNN).

Model Used

• YOLONAS

Model On-boarding

• YOLONAS Model On-boarding Notebook

Android App Source Code

• https://github.com/quic/qidk/tree/master/Solutions/QNN/VisionSolution1-ObjectDetection-YoloNas/app

GenAI Solutions

AI Assistant

Description

AI Assistant is an on-device GenAI chat application that acts as an AI companion, enabling natural language interaction and task assistance. The solution demonstrates how to integrate and accelerate Large Language Models (LLMs) on Snapdragon platforms using GENIE and QAIRT APIs.

Model Used

• LLaMA family of Large Language Models (LLMs)

Model On-boarding

• LLM models are sourced from Qualcomm AI Hub and Hugging Face

• Context binaries are generated using GENIE export flows as described in: LLaMA Model On-boarding from AI Hub

Android App Source Code

• https://github.com/quic/qidk/tree/master/Solutions/GenAI/AI-Assistant/app

Speech to Image

Description

Speech to Image is a multimodal GenAI application that converts spoken input into image outputs fully on-device. The solution demonstrates speech understanding followed by generative image synthesis using Qualcomm GenAI acceleration.

Model Used

• Automatic Speech Recognition model (Speech Encoder)

• Text to Image Generative Model (Diffusion based)

Model On-boarding

• ASR and generative models are downloaded from Qualcomm AI Hub and QPM tutorials

• ASR (Whisper Tiny) and Stable Diffusion Model (QPM tutorials) On-boarding

Android App Source Code

• https://github.com/quic/qidk/tree/master/Solutions/GenAI/speech_to_image/speech_to_image

ASR-LLM-TTS

Description

End-to-end voice AI pipeline that chains Automatic Speech Recognition, a Large Language Model, and Text-to-Speech synthesis entirely on-device. The user speaks a query; the app transcribes it with Whisper (ASR), generates a response with LLaMA (LLM), and speaks the answer back with MeloTTS (TTS) — all accelerated on the Snapdragon NPU via QAIRT GENIE APIs.

Device Compatibility

Device	QAIRT Version	OS
Snapdragon 8 Elite (V79)	2.45.0	Android 15
Snapdragon 8 Elite Gen 5 (V81)	2.45.0	Android 16

Models Used

• Whisper – Automatic Speech Recognition (ASR)

• LLaMA 3.2-3B-Instruct – Large Language Model (LLM)

• MeloTTS – Text-to-Speech (TTS)

Model On-boarding

• ASR (Whisper): Generate via VoiceAI ASR Notebook or download from Qualcomm AI Hub (Whisper)

• LLM (LLaMA): Download optimized LLaMA 3.2-3B model from AI Hub LLM on Genie tutorial or generate via QPM Notebook (LLaMA 3.2-3B IoT)

• TTS (MeloTTS): Generate via VoiceAI TTS Notebook

Android App Source Code

• https://github.com/quic/qidk/tree/master/GenAI-Solutions/ASR-LLM-TTS

VLM – Vision-Language Model

Description

On-device Vision-Language Model (VLM) application that accepts an image and a text question and streams a natural-language answer token-by-token. The solution demonstrates a three-node Genie pipeline — image encoder, text (LUT) encoder, and LLM decoder — running fully on the Hexagon DSP via Qualcomm GENIE and QNN HTP APIs. Supports both image+text and text-only conversation modes.

Device Compatibility

Device	QAIRT Version	OS
Snapdragon 8 Elite Gen 5 (V81)	2.45.0	Android 16

Model Used

• Qwen3-VL-4B-Instruct – Vision-Language Model (w4a16 quantized, QNN HTP)

Model On-boarding

• Download the pre-compiled model bundle (geniex_qairt-w4a16 variant) from Qualcomm AI Hub – Qwen3-VL (Quick Start → OS: Android → Target: Snapdragon 8 Elite Gen 5)

• The required QAIRT SDK version is embedded in metadata.json (tool_versions.qairt) inside the downloaded bundle

Android App Source Code

• https://github.com/quic/qidk/tree/master/GenAI-Solutions/VLM