Ateneo students develop AI-powered software for sari-sari stores

Students of Ateneo de Manila University have developed a software that utilizes artificial intelligence (AI) to assist sari-sari stores and food stalls in analyzing their sales data from handwritten records.

The study by Ateneo's Business Insights Laboratory for Development (BUILD) explores how AI tools, optical character recognition (OCR), and large language model (LLM), can turn handwritten sales logs into manageable digital data.

The software prototype, which was first conceptualized and developed in mid-late 2024, scans logbook photos and uses AI tools to recognize products, match prices, and tabulate sales summaries.

"The research aims to develop a system for translating logbooks into metrics by augmenting manual business processes performed by staff," the abstract reads.

"In technical terms, the pipeline follows a 3-segment approach: an image is uploaded, software tools and AI process and extract the data, and data is tabulated to provide a summary of sales," Zachary Matthew Alabastro, one of the researchers, told GMA News Online.

"AI is integrated to learn from the stall's context and modify the data to output the proper SKU (stock-keeping-units) tags and numbers," he added.

An SKU is a unique alphanumeric code that helps businesses track and manage their inventory.

The study, "Applied Optical Character Recognition and Large Language Models in Augmenting Manual Business Processes for Data Analytics in Traditional Small Businesses with Minimal Digital Adoption",  was published on May 30, 2025. 

It was also presented at the Artificial Intelligence in Human-Computer Interaction Conference 2025, which ran from June 22 to 27, 2025, in Sweden.

Handwriting analysis

Data from a local food stall at the Ateneo's Student Enterprise Center was used for testing, and precision and recall scores were evaluated to examine similarities between original and extracted data.

"Results showed moderate-low precision but above-average recall. The SKUs yielded average precision and recall scores of 0.32 and 0.62, while sales data stood at 0.55 and 0.54, respectively," the abstract read.

The study noted that the software "struggles with accuracy" but has a "fair amount of true values."

"Improvements are needed to enhance the learning mechanism of the LLM," the abstract reads.

"It serves as a potential catalyst for growth by simplifying complex data problems with cost-efficient solutions," it added.

Alabastro said the prototype software is more accessible through a desktop or laptop.

"No further discussions are set, but making it mobile-optimized is on the roadmap," he said.

The software can also be adapted to analyze handwritten data such as inventory sheets, delivery logs, or even payroll ledgers. — VDV, GMA Integrated News