"B Mart carries about 10,000 products, but when customers place an order, they put an average of seven items in the cart. Even if a product is good, if customers don't find it, it won't sell and will eventually disappear. We figured that if we pinpoint only what customers need right now, we could solve both shopping convenience and product exposure at the same time."
On June, the "Nulleobosae" team, made up of non-developers, took third place at the company hackathon (Uathon) of Woowa Brothers, the operator of Baemin and B Mart. The winning entry was "Ssok," an artificial intelligence (AI) shopping agent created by Heo So-yeong, Park Ji-hyun, Jeong Hye-rim and Jeon Da-hae, who handle B Mart's pricing strategy, business management and data analysis. Ssok recommends B Mart products by having AI analyze the meaning and context of a customer's natural-language description of what they want.
It also handles requests with multiple conditions, such as "I live alone and want something spicy for dinner today," "Find me a snack high in protein and low in sugar," and "I have a shrimp allergy; tell me snacks I can eat." It shows the basis for recommendations and nutrition and allergy information right in the search results.
On the 27th, at the Woowa Brothers headquarters in Songpa-gu, Seoul, the Nulleobosae team members said they "got the idea from inconveniences we experienced ourselves." Park said, "For my 30-month-old child with an egg allergy, I had to zoom in on the materials and supplies section of product detail pages to check one by one whether they contain eggs," adding, "Even if I searched for 'food without eggs,' egg cookies would show up, so with existing keyword search alone it was hard to find exactly what I wanted." Park added, "Customers repeat searches about 1.6 times on average before adding a single item to the cart, and we thought if we cut that to once, they could browse more products."
The Nulleobosae team said that although none of the members were developers, their competitive edge came from analyzing customer reactions and business metrics by looking at product and sales data every day. They said they were rated highly not only for technical completeness but also for specifically presenting what problem the service solves and how it could lead to real business outcomes.
Uathon ran for 12 hours, but excluding event briefings, presentations and judging, the actual time available for production was about seven hours. Jeon handled the AI search engine and the allergy information safety mechanism design, and Park was in charge of mapping the customer's product exploration path and user experience and user interface (UX/UI) design. Jeong handled linking B Mart products with Ministery of Food and Drug Safety data and producing the presentation video, and Heo wrote the service proposal and presentation materials and coordinated the overall work. The following is a Q&A.
—What led you to join the hackathon.
(Park Ji-hyun) "As the company actively supported Generative AI tools and vibe coding training, I gained confidence that even non-developers could create real outputs. Previously it was a developer-centered event, but this time we thought we could take on the challenge too."
—Did you plan the AI shopping agent "Ssok" from the beginning.
(Heo So-yeong) "At first, we tried to build a price optimization model that suggests the right discount rate for each product. But the data released a week before the hackathon covered only about 30 days, so we dropped the original idea, which needed at least six months of data. After that, we gathered the inconveniences we experienced while using B Mart and planned a service to help with product discovery."
—How is Ssok different from existing product search.
(Jeon Da-hae) "It doesn't just look up words; it interprets the customer's situation and purpose. If someone says, 'I live alone and want something spicy for dinner today,' it considers small-portion items for one-person households and product categories suitable for dinner. If someone says, 'Grocery shopping for a week,' it even infers the needed quantities and scale."
(Heo So-yeong) "Another differentiator is showing customers the basis for recommendations. AI explains how it interpreted the customer's question and why it recommended these products. If someone is looking for low-sugar snacks, we made it so they can check sugar content right in the search results without going into the product detail page."
—How did you ensure the safety of allergy information.
(Jeon Da-hae) "We separated what AI is good at—semantic inference—from what existing code is good at—making precise determinations. AI broadly finds crustaceans related to shrimp, and the actual safety is reverified with allergy-ingredient data from the Ministery of Food and Drug Safety. AI turns the customer's words into search conditions, and the data makes the final call."
(Jeong Hye-rim) "We also consolidation the Ministery of Food and Drug Safety and B Mart product information. Without barcode information, we cleaned product names to match them, and we brought in more than 5,000 materials and supplies and allergy entries, covering over 70% of food products."
—What were your strengths as a team without developers or designers.
(Jeon Da-hae) "Technically, it was hard to compare ourselves with developer teams. Looking at other teams' screens, we could feel the difference in completeness."
(Heo So-yeong) "Instead, we look at product and sales data every day and analyze customer reactions and business metrics. We presented numbers on how much the number of items per order—an average of seven—could increase if we reduced the average 1.6 searches customers make before adding a single item to just one, and what impact that would have on contribution margin. Even as non-developers, our strength was explaining the problem in connection with business outcomes."
(Park Ji-hyun) "I handled UX/UI design on the team, and even with good ideas and technology, if they aren't conveyed intuitively to customers, the practical utility can fall short. We put a lot of thought into how to show the information customers care about most—like allergies and nutrition—simply and prominently."
—How much actual production time did you have.
(Jeong Hye-rim) "The event ran for 12 hours, but the actual working time was about seven hours, from 11 a.m. to 6 p.m. We finished data linking, service implementation, drafting the proposal and preparing the presentation within that window. We also made a 90-second demo video using AI voice, and we ranked first in a peer review that determined who advanced to the finals."
—What needs to be improved to launch this as a real B Mart service.
(Jeon Da-hae) "We need to solve the expense and system load that occur when countless customers use AI search simultaneously. Another task is how accurately it reflects B Mart customer purchasing tendencies."
(Jeong Hye-rim) "There are cases where Ministery of Food and Drug Safety information doesn't match the actual product information, and many product details are registered as images. Because allergies are sensitive information, precise extraction and verification of data must come first."
—What did you take away from this experience.
(Heo So-yeong) "Defining what the problem is will likely become more important going forward. Implementation can get help from AI, but discovering problems in the field and deciding why they should be improved is the role of people."
(Park Ji-hyun) "In the past, even if I had ideas, I felt limited in actually implementing them due to a lack of technical knowledge. Now, if what I want to do is clear and the problems I want to solve at work are specific, I feel confident I can use AI to build deliverables at a considerable level."