Deep Learning Architectures for an Integrated AI System in Retail: A Case Study of Adepa SuperMart, Accra

Artificial intelligence has transformed retail by enabling automated quality inspection, data-driven demand forecasting, and always-available customer support. This paper presents a concrete, end-to-end case study of an integrated deep learning system designed for Adepa SuperMart, a fictional supermarket in Accra, Ghana, selling groceries including canned and boxed goods. The paper addresses three components of such a system: a Convolutional Neural Network (CNN) for detecting bulging or swollen cans and boxes, a food-safety indicator well documented in the food science literature; a Recurrent Neural Network (RNN, specifically an LSTM) for forecasting daily sales from historical transaction data; and a Large Language Model (LLM)-based chatbot, deployable entirely on local infrastructure using the open-source Ollama runtime, for customer support. Each component is implemented in Python and evaluated on real or realistically grounded data: the CNN is trained on a programmatically generated synthetic image set (disclosed in full, since a real photographed defect dataset was not accessible from this environment) and achieves 100% test accuracy, a result discussed critically as a synthetic-data ceiling effect rather than a real-world performance claim; the RNN is trained on real point-of-sale transaction data (1,000 real transactions from a public retail dataset) combined with a literature-grounded synthetic overlay representing the effect of visible product defects on purchase behaviour, and achieves a modest, honestly reported R² of 0.007 on a held-out test set, a result used to illustrate the genuine data-volume requirements of deep learning in a data-scarce retail setting. The paper situates all three components within a full data engineering lifecycle, spanning data cleaning, model training, deployment, operations and maintenance, and ethics and governance, and closes with a critical assessment of what these results do and do not demonstrate.