Wiring Android's NNAPI to a Quantized On-Device LLM for Real-Time Function Calling

Deep technical walkthrough of running structured output generation (function calling / tool use) from a quantized LLM on Android using NNAPI delegation — covering JSON schema injection into the prompt prefix, enforced token budgets to guarantee sub-300ms tool-call latency, output parsing with fallback heuristics when the model drifts from the schema, and the memory/compute tradeoffs that determine whether a 3B or 7B model fits your target device tier