I already have an Android app written in Kotlin that upscales images with TensorFlow Lite. Right now the model runs on CPU and GPU; I’d like a concise upgrade that lets it execute on the Neural Processing Unit found in recent Qualcomm Snapdragon chips. Key details • Tech stack in place: Kotlin + TensorFlow Lite. • Target hardware: recent Snapdragon devices (e.g., 855, 865, 888—exact list can be refined together). • Priority: unlock NPU performance while preserving the current accuracy of results. What I need from you 1. Modify the existing TensorFlow Lite setup so inference routes through the Snapdragon NPU (likely via the Qualcomm Neural Processing SDK / QNN delegate or any newer, officially supported path). 2. Keep model output numerically in line with current CPU/GPU results; any deviation must be explained and, if practical, corrected. 3. Supply minimal but complete deliverables: • Updated Kotlin source showing the delegate switch. • Gradle / CMake tweaks, if required, for SDK integration. • A short README with build steps, device prerequisites, and a quick benchmark script or log showcasing the speed gain. I’m unsure whether the relevant Qualcomm libraries are already bundled, so part of the task is advising on—or fetching—what’s necessary and documenting it clearly. Scope is intentionally focused: just the delegate integration plus proof it works and still meets accuracy expectations. If that goes smoothly there may be follow-on work for multi-model support and broader device coverage. You can check app here for reference :- https://indusapp.store/db7nthjj