[Interview] Shin Jae-hoon, Head of the Agricultural Intelligence Data Team at the Rural Development Administration: “Isaac does what general-purpose AI cannot”

‘AI Isaac’ spent its first year after launch quietly. After being unveiled in December 2024, it did not have many users. The number of questions received in December 2025 was 2,542. This year, from January to September alone, 104,702 questions accumulated.

The mood changed this year. From January, monthly inquiries exceeded 10,000, and on February 19, a dedicated organization, the Agricultural Intelligence Data Team, was created at the Rural Development Administration. Seven months later, on the 30th of last month, the team released ‘AI Isaac 2.0’. It is a ‘farming secretary’ that diagnoses pests and diseases from photos, prescribes fertilizer amounts based on soil test results, and even understands dialect questions.

On the 8th, we met team leader Shin Jae-hoon, who has led the team since its launch. He began by saying, “I want to clear up the misunderstanding of whether Isaac is needed when ChatGPT is this good.”
 

From 2,542 questions in the first year after launch to the creation of a dedicated team

-How did AI Isaac begin?
“ChatGPT came out at the end of November 2022. We had long thought a chatbot was needed for farming consultations, but a simple chatbot had limitations. Since this technology had emerged, we wanted to use it. When we looked around, there was a domestic model, Naver HyperCLOVA X. It had the advantage of being good at Korean because it had learned Korean. Just then, the Ministry of Science and ICT put out a call for projects using domestic models, and together with Naver and an AI-specialized company, we proposed an agricultural technology consultation chatbot and were selected in 2024. Development began that year and it was opened to the public in December.”
 

-What was the initial response?
“It is difficult for farmers to enter an information system and search, but it was effective because they could ask verbally and receive answers. However, as ChatGPT continued to be upgraded afterward, our answers looked relatively plain. It was not easy to activate usage, and last year passed with it known only to the extent of ‘there is something like this.’”
 

-What prompted the renewed momentum?
“At the end of last year, the government put forward the goal of becoming one of the ‘top three AI powers’. Commissioner Lee Seung-don personally briefed the ‘Agricultural Science and Technology Artificial Intelligence (AI) Convergence Strategy’ last November and named the provision of agricultural technology using Isaac as the No. 1 task. The goal was to create an app that farmers would ask questions of every day and receive briefings from. That is how the dedicated team was created on February 19 this year. I had experience planning digital agriculture as head of the data team in 2021~2022, and I gathered all the people who had worked with me then. The staff consists of 11 people, including IT personnel, agricultural researchers, and rural extension officers.”

Into an AI agent that calls up pest and disease diagnosis and fertilizer prescriptions

-What changed the most in 2.0?
“The core of 1.0 was to find content similar to a question from 560,000 pages of materials created by the RDA and answer based on it. But ChatGPT also searches public data and answers well. Because we answer with verified materials, it is trustworthy, but there was no function that general-purpose AI could not do. So in 2.0, Isaac calls up and answers with specialized functions that the RDA has in separate systems, such as pest and disease diagnosis, fertilizer prescriptions, and weather disaster prediction (early warning is scheduled to be connected by the end of the year). Isaac does what general-purpose AI cannot.”

Field-photo accuracy from 50~60% to 92%

-In the RDA’s own comparison, GPT-5.6 correctly identified 46.5% of pest and disease photos, while Isaac got 92% right. What was hardest in raising it to 92%?
“During research, diagnosis with photos collected by researchers yielded 90~95%. But photos taken by farmers in the field were not like that. When we tried ordinary photos, it was 50~60%. The commissioner, a pest and disease expert, said we could not put this into Isaac as it was, so we worked again through the nights for two months. We inspected all 300,000 photos and raised it to 92%. We collected representative symptoms photographed directly by researchers, ambiguous early-stage photos, and normal photos for comparison, and for insect pests we trained the model on the entire process from larvae and nymphs to adults. It is work only experts can do. Because we had already done the first-stage project for pest and disease diagnosis, we could immediately integrate it into Isaac.”
 

-What about cases such as mixed infections that are difficult even for people to distinguish?
“What people cannot do, AI cannot do either. That is because artificial intelligence learns by following people. The target crops are also still insufficient. The 25 crops that can currently be diagnosed account for about 74% of cultivated area, so the remaining 26% cannot be diagnosed. We are working to expand to 60 crops by 2029 so that it can cover more than 90%.”

 Summarizing only registered documents and comparing with 300 answer keys

-How do you prevent hallucinations in generative AI?
“General-purpose AI learns vast amounts of data and answers based on it, so incorrect information can be included, and if it does not know, it may make things up. That is because it has been trained to answer. By contrast, Isaac searches registered documents, selects documents related to the question, summarizes them, and answers. It uses no external materials at all. Performance is about finding relevant content well, and the key is summarizing it in an easy-to-understand way. System information such as Heuktōram (the RDA Soil Environment Information System) is also already verified information, so you can see it as minimizing hallucinations.”
 

-The RDA formed a quality assurance group on the 1st of last month. How do you manage answer quality?
“We have five members of the Isaac team and one employee dedicated to quality, and we continuously monitor answers. If an unsatisfactory answer appears, we find related documents and register them again. Twenty-five technical committee members from the Customer Support Center also participate. They are retirees or experts in the relevant fields, and they receive complaints and, when necessary, go to the field to provide technical guidance. We compare Isaac’s answers with 300 answer keys created by these people and fill in shortcomings. It is not a matter of making it and saying, ‘use it if you like it, and don’t if you don’t’; we have established a system for continuous management. As the number of users grows, problems we had not experienced before are also appearing, and we are taking action immediately as they arise.”
 

Letting 250,000 participants in New Year practical farmer training experience Isaac

-Can elderly farmers use it too?
“Seeing it at the National Convention of Korean Rural Leaders held in Cheonan from the 6th to the 8th, we judged that people in their 60s and 70s could use it sufficiently. If they can use a smartphone to some extent, they only need to press buttons. There was feedback that the text in 1.0 was too small, so in 2.0 we made the text larger as well. However, for now, we think young farmers and returning farmers will approach it more easily. They do not know farming well, so they will search anyway, but general-purpose AI can provide incorrect information. The idea is to search, but verify once more with Isaac.”
 

-What are your plans to expand Isaac?
“Education is important. Agricultural technology center curricula include 10,000 people in farmer colleges, 30,000 in agricultural machinery training, and 250,000 in New Year practical farmer training, and we plan to introduce Isaac without fail there and have participants experience it directly. We conducted first-stage training for 120 farmers belonging to the Korea Information-oriented Farmers’ Central Association as instructors. They will inform their peers and also work as instructors for New Year practical farmer training. We are also working to connect Isaac to Daedong’s ‘AI Daedong-i’, so AI Daedong-i answers questions on agricultural machinery maintenance and Isaac answers farming-technology questions, and we aim to make Isaac available on public and private platforms such as NongHyup’s ‘NH Today’s Farming’.”
 

The core of 3.0 is farm income… Next year’s budget of 1.3 billion won reflected in the government proposal

-What will be included in 3.0, and when will it come out?
“First is income. Using Isaac should help farms, and the basis for raising income begins with diagnosis. There is already a management diagnosis function, but we want to advance it to find areas for improvement and strengthen management consulting that proposes agricultural technologies suited to them. Second, because it currently only searches existing documents and answers, it cannot update documents immediately, so the latest information is weak. Next year, we plan to add a web search function, but because it must not search irrelevant places, it will search and summarize only authorized sources. Legal information will be attached in November, and policy information will also be added. The language model we are using now is a 2024 model, which is somewhat disappointing, but if we apply the national representative AI model the government is creating, the data will remain the same while the explanations will improve. 3.0 will be released next year, and before then we will add good functions from time to time.”
 

-Is the budget sufficient?
“This year’s budget was 100 million won. For next year, 1.3 billion won has been reflected in the government proposal. It has not yet passed the National Assembly, but if we secure enough formal budget, farmers will be able to use it much better.”
 

-Any final words?
“This is something many farmers misunderstand: general-purpose AI has various functions, but when it comes to farming technology, Isaac is the best. It gives accurate and reliable answers. Isaac is also included in the Ministry of the Interior and Safety’s ‘30 Key Tasks for Realizing an AI Democratic Government’. The structure is that general administration is handled by the AI national secretary, and agricultural technology by Isaac. The goal is clear. Farmers should ask Isaac every day and receive briefings every morning.”