Multimodal AI Food Retrieval on Qualcomm Dragonwing IQ-9075

Multimodal food understanding is becoming increasingly relevant for edge AI applications, enabling intelligent assistance for nutrition, food recognition, dietary awareness, smart kitchens, and automated food-service systems. In this project, we developed an on-device multimodal food retrieval pipeline for the Qualcomm Dragonwing IQ-9075 EVK, combining image-text retrieval with vision-language reasoning to recognize ingredients, retrieve detailed dish descriptions, and estimate calories directly from food images. The system uses SigLIP2 for efficient candidate retrieval, Qwen3-VL for closed-choice ingredient selection, and a compact Qwen3-VL reranker for fine-grained caption matching. Confidence-based execution and model-specific quantization reduce unnecessary VLM computation, improving the trade-off between accuracy and efficiency. All inference is performed locally on the device, demonstrating how retrieval and multimodal reasoning can be combined in a hardware-accelerated pipeline for practical edge AI deployment.