We built an enterprise-grade expense tracking solution by fine-tuning the Gemma 4 E2B vision model using Unsloth LoRA on the ICDAR 2019 SROIE dataset. While standard zero-shot models hallucinate and fail at strict JSON formatting, our model accurately extracts merchants, dates, addresses, and totals from chaotic receipt layouts. We deployed this pipeline into an interactive Gradio dashboard featuring a RAG-based vendor matching engine to auto-retrieve tax IDs, multi-receipt batch processing with CSV exports, and one-click Hugging Face model publishing. Our solution transforms raw receipt images into production-ready structured data.