Used only where a person used to decide
We don't use AI for work that can be written as rules. We use it only where rules grow long and exceptions keep appearing — where a person used to look and decide every time.
Public cases 4 · 2024–2026
Attaching AI is not the goal.
Putting AI into work a rule can handle makes it slower and costlier — and the same input gives different answers. Conversely, forcing rules to hold a spot where exceptions keep appearing makes the rules more complex than the person. We first tell the two apart.
-
A person reviews it last
What AI produces is a draft. It isn't sent out as is; a person checks it and then confirms it. The moment it goes out automatically, there is no room left to take it back when it is wrong.
-
It can be traced back
It must be possible to go back from the result to the source data. If you don't know which sentence a value came from, you can't find what went wrong when it does.
-
We don't lean on the model's name
Models change. In one of our products, the model actually got swapped during operation. We build it so a new one fits into the same spot.
-
We don't turn what we haven't measured into numbers
We don't casually write 'accuracy of X percent.' We publish the figure only for cases with a measurement record, and where there is none, we say so.
What we built for you.
Each case shows the situation before it was commissioned, the result, and the actual screens. Some were carried out as national R&D projects.
-
A document that arrives by email becomes a row in Excel, with no human hand
It watches the inbox, reads the text out of attached PDFs and images, and an AI splits that text into fields and delivers it to Excel. The form is defined by whoever sends it.
Service in operation Deployed 2026.01 from Pilot service · running free of charge to
-
Reviews from three platforms gathered on one screen, and the AI writes the replies first
Reviews from three delivery platforms are gathered automatically, and the AI drafts the replies. The owner reads them and just presses confirm to post.
Review replies Written by hand across three sites from AI draft, posted after confirmation to
-
An analysis tool that shows thousands of research comments split by topic and sentiment
Video comments gathered by search term are grouped by topic, with a positive/negative distribution alongside. How many topics to split into is decided by the researcher on screen. It was built so the researcher can change the criteria and re-run it themselves.
Understanding comment topics By skimming with the eye and guessing from By per-topic, per-sentiment distribution to
-
The forgery verdict comes at the very moment identity verification is in progress
Technology that analyzes forgery after the fact already exists. If the verdict comes too late, there is nothing you can do with it, so the goal was to get the verdict while the procedure is still in progress. It is a national R&D project carried out under the 중소벤처기업부 (Ministry of SMEs and Startups) 창업성장기술개발사업 (Startup Growth Technology Development Program), and the outcome was AntiFake, for which a patent was filed.
Project execution Started 2023.05 from Completed 2024.07 · all 8 performance items achieved to
Frequently asked about this topic
If data is needed before the analysis.
For AI to decide, data to decide on must come first. Gathering data is Web data collection and App crawling, and counting what was gathered by the same criteria is Data analysis .
We start by checking whether this is a place for AI.
Tell us about a procedure you currently decide on by eye every time, and we'll first sort out whether it is work a rule can handle or work that needs AI.
Replies within 1 business day