Developing robust conversational agents for e-commerce, particularly for voice-activated product ordering, presents a significant challenge. This paper introduces a lightweight and efficient framework for this task, centered on a compact ( \(\sim \) 1B parameter) Large Language Model (LLM), Gemma-3 1B. The proposed methodology leverages parameter-efficient fine-tuning (PEFT) through Low-Rank Adaptation (LoRA) to specialize the model for semantic parsing, combined with post-training quantization to ensure efficient deployment on both server and mobile platforms. The system is trained on a synthetic dataset of 20,000 spoken shopping requests, enabling it to map diverse user utterances to precise JSON order specifications. Experimental evaluation demonstrates the high efficacy of this approach. The LoRA-tuned model achieves an exact JSON match accuracy of approximately 94.6% and a slot-level F1 score of nearly 97%, demonstrating performance that is competitive with full fine-tuning. Furthermore, the system exhibits practical deployment characteristics, with sub-second latency on server infrastructure ( \(\sim \) 0.25 s) and feasible on-device latency ( \(\sim \) 2.5 s) when using 8-bit quantization. These findings demonstrate that compact LLMs, when coupled with LoRA-based specialization and quantization, can serve as both accurate and deployable semantic parsers for voice-based commerce applications. The proposed framework not only achieves competitive accuracy with significantly reduced hardware requirements but also establishes a reproducible blueprint for real-world conversational AI deployment.