How Big-Tech Companies Like Naver, Kakao, and Toss Use AI

We've put together how Korea's leading tech startups are applying generative AI.

We recommend this article for

✅ Anyone who's wondered whether AI can really be applied to their work

✅ Anyone who's been looking for areas they can automate with AI

A summary of how big-tech companies use AI

• Gangnam Unni: Creating marketing images using AI models

• Baemin: Menu recommendations through review summaries, an in-house AI data analyst

• Toss: Creating icons and content graphic images

• Kurly: Improving search results, automating review analysis

• Naver: Developer code reviews

• Kakao: Spam-content classification

How Gangnam Unni uses AI

Creating marketing images with AI

• Description: They use AI-generated models in their marketing materials.

• Model used: Midjourney

• Background: Finding models wasn't easy. In an industry that deals with cosmetic surgery and procedures, models tend to be less willing because the work requires photographing various body parts. Hair, makeup, and wardrobe also cost a lot. And whenever a contract expired, they had to find a new model and replace the images.

• Use cases

1) You can dress AI models in a variety of outfits.

2) You can also create different moods depending on hair length or color.

3) And you can easily change the background.

• User feedback

1) "It's great that you can freely generate the model's poses, hairstyles, and clothes—so much so that people say, in professional terms, the look has really improved."

2) "There's no need to worry about portrait rights the way you would with a celebrity, so it's much safer."

3) "As we expand globally, we can localize and use model images tailored to each country."

How Baemin uses AI

Menu recommendations through review summaries

• Description: AI summarizes review data to extract key keywords, then uses them to recommend menus.

• Model used: GPT-4

• Background: They discovered the insight that a significant number of Baemin app users open the app without having decided on a menu or restaurant in advance. That led them to the idea that GPT could do a good job of recommending menus and restaurants.

In-house AI data analyst

• Description: They built an AI data analyst that explains the database, writes SQL queries, and explains the analysis results.

Baemin employee: "As of July 1, 2024, tell me the number of delivery orders for XXX Food delivery, single servings, alcohol, noodles, and coffee menus, along with each one's ratio relative to total food delivery orders."

AI: "This SQL query calculates the data you described as of July 1, 2024."

• Models used: GPT-3.5, GPT-4o

• Background: You can write queries using GPT-4 alone, but it had limitations for real-world work because it lacked an understanding of the company's internal domain and data policies. To overcome this, Baemin used Langchain.

• User feedback: "It's especially helpful when you've just joined the company or taken on work in a different domain."

How Toss uses AI

AI graphic generator

• Description: They built a tool that lets AI generate the graphic images used in the Toss app.

• Model used: Midjourney

• Background: They needed graphic images for everything from app icons to content images, but there weren't enough graphic designers.

• Use case: They can now create graphics without having to ask a graphic designer, which sped up their experimentation.

How Kurly uses AI

Improving the search engine

• Description: They used an LLM to improve search-result performance.

• Model used: Google Cloud Vertex AI Search

• Background: Cases where search results couldn't be provided due to typos, spacing mistakes, and the like accounted for about 6% of all searches. In those cases, they were showing best-selling products. To show similar results for cases the existing search engine couldn't handle, they needed technology that could understand context well.

• Use case: The cases where no search results could be provided dropped significantly, and the customer cart-conversion rate and cart value both went up.

Automating review-analysis work

• Description: They built a process that uses AI to classify and analyze countless reviews.

• Model used: Gemini

• Use cases

1) Classifying reviews by category

2) Analyzing the customer sentiment expressed in reviews

How Naver uses AI

AI code review using an LLM

Description: AI looks at the code a developer wrote and leaves feedback on points to improve.

AI: "Using thread sleep to implement retries is not a good idea. Thread sleep can cause unpredictable delays. Instead, you can implement retries using XXX's after."

AI: "There are cases where it returns null. That's a bad practice. It's better to use Optional or throw an exception."

• Model used: Llama3

How Kakao uses AI

An LLM for handling spam content

• Model used: developed in-house

• Background: Rather than using an existing LLM, they built their own model. They say this was because the model needed to know Kakao's regulatory policies, have no security issues, and not be cost-prohibitive. Also, since there wasn't much data for distinguishing spam, they hired labelers and did additional work labeling spam data along with the reasons.

• Description: They built an AI model that finds sentences it considers spam and even identifies why they're spam.

• Model used: developed in-house

• Background: Rather than using an existing LLM, they built their own model. They say this was because the model needed to know Kakao's regulatory policies, have no security issues, and not be cost-prohibitive. Also, since there wasn't much data for distinguishing spam, they hired labelers and did additional work labeling spam data along with the reasons.

To sum up once more…

Here's how they're using generative AI.

• Gangnam Unni: Creating marketing images

• Baemin: Menu recommendations through review summaries, an in-house AI data analyst

• Toss: Creating icons and content graphic images

• Kurly: Improving search results, automating review analysis

• Naver: Developer code reviews

• Kakao: Spam-content classification

At Dalpha, where AI experts come together, we offer free AI consultations.

At Dalpha, through 1:1 consultations,
custom AI solutions tailored to each company and product—we plan and build them for you.

If you're curious about a free consultation, feel free to reach out anytime!

Apply for a free AI consultation

References (sources)

• Gangnam Unni (image generation): https://blog.gangnamunni.com/post/GenerativeAI/

• Baemin (review summaries, recommendations): https://techblog.woowahan.com/16877/

• Baemin (data analysis): https://techblog.woowahan.com/18144/

• Toss (image generation): https://toss.tech/article/ai-graphic-generator-2

• Kurly (search): https://helloworld.kurly.com/blog/vertex-ai-search-NR/

• Kurly (review analysis): https://helloworld.kurly.com/blog/bigquery-gemini-review/

• Naver (code review): https://d2.naver.com/helloworld/7321313

• Kakao (spam classification): https://www.youtube.com/watch?v=W0WvXVlaPS4

Articles worth reading together 📝

• 5 AI solutions frequently used by content/media companies
  

• Enterprise AI in action: 7 cases (feat. Danggeun Market (Karrot), Dropbox, Geek News)

• 'Dingo (Makeus)' cut short-form production costs by 35% by adopting short-form generation AI!

• The 10-step guide to AI development—everything you need in this one article