Inspiration
South Korea has grown rapidly. A country that grew rice amid the ruins of war now also produces semiconductors, often called the “rice of industry.” Yet the people who have sustained its agriculture are also aging rapidly.
According to Korea’s 2025 Agriculture, Forestry and Fisheries Census, 51.3% of the farm-household population was aged 65 or older, and 61.8% of farming households cultivated less than 0.5 hectares. As older farmers find it increasingly difficult to manage these small holdings on their own, who will carry farming forward—and how? Official 2025 census results
Korea’s Farmland Act generally ties farmland ownership to the owner’s agricultural management, while defining exceptions to ownership restrictions and the circumstances in which entrusted farming is permitted. Keeping agriculture viable therefore requires attention to institutions and operating arrangements alongside technology. If new technologies are shaping a new era, could they also change the way we farm? Farmland Act, Article 6, Article 9
K-Farm began with that question. It is an agricultural operations system and a policy proposal to help Korea’s aging rural communities keep farming efficiently. The proposal connects dispersed fields, workers and equipment through accessible local centers, with AI handling crop selection and routine planning.
At its center is one question: what should be grown on each parcel, and how should crop choices across multiple parcels be planned together? K-Farm aims to consider land and climate suitability, regional characteristics, prices and supply conditions so that individual crop choices contribute to a coordinated regional cultivation plan.
Scheduling, labor and equipment management turn that crop plan into farm work. Centers can share equipment with neighboring centers or consider acquiring what they need, while connecting workers with tasks at the right time. The center manager acts as an intermediary who understands landowners’ and workers’ circumstances and uses AI assistance to resolve exceptions and errors.
When entrusted farming continues across seasons, each parcel accumulates cultivation experience and work records. Those records could help centers choose crops better suited to their region, prepare equipment and labor for recurring cultivation, and gradually coordinate crop composition across participating regions. I believe that cooperation, sustained over time, could provide a foundation for improving how Korean agriculture operates as a whole.
This policy proposal builds on an existing foundation. Korea already has institutions and programs such as Nonghyup agricultural cooperatives, the Farmland Bank, farm-work services and rural labor support. K-Farm proposes an operating approach that connects that foundation with new technology. I am developing it as working software so that the institutional review and field validation needed for real adoption can be discussed in concrete terms. MAFRA: Farmland Bank lease entrustment, MAFRA: agricultural support program guide
Through this hackathon, I hope to create an opportunity for K-Farm to advance into real policy discussions and field application. Keeping land in cultivation as communities age, bringing local people and resources together, and finding a sustainable way forward for Korean agriculture are what inspired this project.
What it does
K-Farm is an AI agricultural operations system that helps local centers choose crops for entrusted farmland and turn cultivation plans into day-to-day work. AI develops routine plans and proposed changes, while center staff connect landowners and workers and manage execution and exceptions.
1. Register farmland and cultivation conditions.
Center staff register applicants and parcels, check cadastral maps and review field conditions. New applications can be incorporated as they arrive while preserving existing cultivation plans and completed work.
2. Compare crop options and crop composition across multiple fields.
NVIDIA Nemotron, running through Nebius Token Factory, proposes crop combinations based on field conditions and cultivation criteria. It uses a calculation tool to compare alternatives, assess their work and resource requirements, and revise its selection when a proposed plan fails validation. The review considers the current group of fields and the remaining registered fields with known crop options. Where reviewed evidence is configured, participant production limits, regional varieties, prices, costs and available working hours also inform the comparison. Missing evidence remains a condition requiring review.
Price, cost and climate references are stored with their sources and reference dates. Staff can inspect the evidence supporting crop decisions within the same application.
3. Translate crop plans into schedules and resource requirements.
The system calculates work periods, equipment requirements and labor demand for planting, crop care and harvesting. Registered workers are distinguished from workers with confirmed working calendars, including qualified equipment operators. Plans that pass the staff execution or approval workflow are saved together as cultivation records, work schedules, equipment reservations and labor requirements. Unsecured resources remain outstanding needs. Equipment shortages can leave a review record for scheduling, neighboring-center capacity, rental or acquisition; the system does not authorize a purchase.
4. Adapt plans when problems arise during cultivation.
When staff report equipment failure, delayed work or worker absence, the system reassesses schedules and resources to carry out the existing crop plan. It considers available replacement equipment and preauthorized resources from neighboring centers, then saves validated changes with their history. Completed work and other centers’ confirmed commitments are preserved.
5. Manage local workers independently of land entrustment.
People can register for work regardless of whether they own farmland. Staff manage the number of workers needed for each task and confirm attendance. The text-message screen manages queued notices, attendance status and resend requests. Actual SMS delivery and wage payments require separate integrations.
6. Maintain each parcel’s cultivation and work history.
Staff can inspect current and next crops, upcoming tasks, recorded work, equipment use, attendance and changes for each parcel. New cultivation cycles retain links to earlier cycles and preserve completed work. The next season requires reviewed cultivation criteria rather than automatically treating an old plan as valid. Where enough comparable, confirmed manual-work records are available, the calculation can adjust workload estimates while keeping simulated records separate from actual field results.
7. Show crop plans and progress across participating regions.
The national overview presents participating centers’ crop areas, expected output and progress by region and period. Township staff can see what is planned beyond their own area. The totals cover centers and farmland registered in the system.
K-Farm is being developed for real use. Demonstrations enter prepared field, worker and equipment data into the same application and follow its registration, AI planning, validation and storage workflow.
Try the English judge environment: a restricted account is prefilled; click Sign in, then use Schedule → 2027 → May/October for saved work and National overview → 2027 for the participating network. The judge guide separates inspection from actions that save new plans.
How we built it
K-Farm was developed so that land registration, crop selection, cultivation planning, work changes and record storage form one connected workflow. The web interface uses Next.js, React and TypeScript. PostgreSQL stores farmland, cultivation, work and reservation records, with Prisma connecting the application to the database.
For crop planning, we use NVIDIA Nemotron through Nebius Token Factory. The model proposes crop combinations and resource-use scenarios, then calls evaluate_plan to check the required work and feasibility. A bounded search can supply additional calculated combinations when the initial group cannot meet its constraints. The model must assess the evaluated alternatives before choosing a plan with select_plan. The server grounds the evidence in calculated metrics and rejects unsupported, infeasible or dominated selections. A bounded correction opportunity does not relax those checks.
A Python service using OR-Tools CP-SAT handles calculation and schedule validation. It checks crop-specific work windows, task order, equipment, qualified operators and available working hours. Configured participant production limits are checked against their stated scope and evidence. Variety checks can account for reviewed seed or trial-area limits and existing orchard age and pollination conditions. These controls support evidence-based planning for participating fields; they do not infer Korea's adequate production level from the number of registered parcels.
When staff execute or approve a plan, the system rechecks permissions, record changes and resource availability. Plans that pass validation are applied together to cultivation records, work schedules and reservations. If a connection is interrupted or a request is repeated, the stored result can be recovered. Actual model usage and costs are recorded in a separate ledger.
VWorld provides cadastral lookup, and Korea Meteorological Administration data supports weather information. Price, cost and other references are stored with their sources and reference dates so staff and AI have evidence to review. Service API keys are managed on the server.
During development, I used OpenAI Codex for coding, debugging and testing. Using the application's operating workflow with separate validation databases, I checked actual-model plan application, persistence after reconnecting and preservation of existing records. The published video was recorded in the public application on October 28, 2026. It shows saved application records and separately dated October 27 AI and data-handling validation evidence; no new model call or approval was performed during filming.
Validation results
K-Farm was tested through actual AI model proposals, calculation and validation, staff execution or approval, and persistent records. The latest checks focused on whether crop composition responds usefully to constraints and whether entrusted fields can continue into another season without resetting their history.
Crop composition under shared constraints: In one prespecified synthetic case covering the same 3,100 m², the earlier local AI adviser selected pepper throughout and required 74.5 calculated person-hours. The improved candidate selected 1,600 m² of potato and 1,500 m² of pepper, requiring 52 person-hours: 30.2% less calculated labor in that case, with no missing essential tasks or production-limit violations. A second case was unchanged. In a third, the earlier selection was rejected and the improved candidate produced a valid plan. These are three bounded comparisons, not an average farm-efficiency gain or proof of nationwide optimality.
Continuity from 2026 to 2027: Three normally registered synthetic fields completed actual-model planning and storage for 2026. Seventeen scheduled tasks were recorded as simulated completions. Without next-season criteria, all three fields were held. After reviewed synthetic 2027 criteria were supplied, actual-model planning created and saved the next season's plans. Six cultivation records retained three links to their preceding cycles, with no overlapping cycles or duplicate task creation. Earlier records and the existing load dataset remained intact. This checks the operating workflow across seasons, not a completed real farming season.
Handling accumulated records: A separate load test in the retained local database stored 50,000 synthetic draft cultivation records and compiled 572,784 task drafts. After correcting an oversized database read, aggregation completed in 22.8 seconds on the tested environment. These were data-handling checks, not 50,000 model calls or approved plans. Production-yield references were missing for those load records, so their count does not establish national production coverage or supply balance.
The final candidate was checked separately using recorded-response regression and actual-model annual planning. Versioned evidence distinguishes these results from the earlier comparisons and hosted and UI checks. These synthetic software tests do not establish real harvest results, actual profitability or nationwide optimal crop allocation.
AI decision evidence · Verification record
Challenges and what we learned
The hardest challenge was defining K-Farm’s operating model itself. The system I wanted to build had no settled form. Where should K-Farm begin, and how far should its responsibilities extend? Could existing policies be connected, or would new institutional arrangements and support be needed? Without clear boundaries, it was difficult to decide how much of the software to implement.
Registering land, choosing crops and assigning work did not seem especially difficult when considered separately. I wanted to catalog every crop grown in Korea and connect planting, crop care and harvesting to suitable equipment and workers. I wanted to reflect regional climates and specialties, demonstrate the system nationwide and eventually use technologies such as autonomous tractors to keep cultivation going with minimal labor.
Including policy made the scope wider still. From finding an institutional pathway for entrusted farming to disaster-loss support, machinery rental and support for older people participating in agricultural work, I wanted to connect existing programs and future support within one operating structure. The common purpose was to help agriculture continue.
But I could not collect and validate all the evidence this vision required within the available time. The institutional and support arrangements also needed separate discussion and review. For this development, I assumed that the necessary institutional conditions could be established and focused on testing the software’s operating workflow under those assumptions. I retained a 30-crop catalog while distinguishing the uneven evidence available for each crop, and narrowed the demonstration to four centers.
Turning AI judgments into executable plans was another challenge. Even when data was available, a plausible explanation was not enough to establish feasibility. The model could cite an incorrect figure or rely on equipment and workers that were no longer available. Optimal nationwide crop allocation remained a goal, not a result I could claim.
This led me to separate the staff-facing interface from the detailed decision evidence. Staff see understandable plans and records, while the system retains the inputs, calculations and reasons behind AI decisions. Conditions are checked again immediately before a plan is applied.
During development, I found myself wondering: “Farmland, crops, schedules, equipment, workers—can I really build a system this large on my own?” Each feature created more connections to work through, and I began to understand how much structure the original idea required.
But as I broke that structure down and implemented it with AI assistance, I watched an idea in my head become a working application. This project changed what I thought I could attempt on my own. There is still much to resolve, but I learned that one person’s idea can become a system others can explore and use as a starting point for discussing what a policy could make possible.
What's next
K-Farm's next goal is to find an opportunity to test this operating model through a pilot in one local area. I would like to work with local landowners, agricultural practitioners and a prospective center operator to agree on participation and operating conditions, then adapt K-Farm to that area's land, people and farming needs.
The pilot would follow a full growing season, from crop selection and cultivation planning through coordinating workers and equipment, responding to changes and harvesting. I want to learn how much routine planning AI can handle, when center staff need to intervene, and what landowners and workers need to remain involved. Price, cost and climate references would be refined for the pilot area, with plans compared against actual farming outcomes to improve the decision criteria.
The software would also develop toward making the fullest practical use of agricultural automation, including autonomous tractors and drones. Based on the land and crop conditions, it would distinguish work machines can perform from work that requires people. The pilot would test integrations that send work plans to compatible equipment and receive execution records in return. The aim is to put machines to work on repetitive, physically demanding tasks and concentrate human effort on field judgment and work that machines cannot readily replace, helping aging communities sustain agriculture with fewer available workers.
I also want to complete the connections needed for day-to-day operation. Actual SMS delivery and incoming replies would connect to the existing labor and messaging workflows, supporting attendance confirmation and change notices. Work, harvest, shipment and actual cost records would feed into financial settlement. Wage payments would be managed separately from crop-sale proceeds, so the center could pay workers before harvest and account for those costs later when settling with landowners. Following a successful harvest and sale, the records would show how the center's agreed remuneration and the landowner's proceeds were calculated.
The software would distinguish disaster losses, losses attributable to center operations and owner-initiated cancellation. Disaster records would support assessment of applicable insurance and public assistance, while losses attributable to the center would retain evidence for compensation decisions. If an owner cancels the arrangement, the workflow would support reimbursement to the center for documented costs already incurred, including labor and equipment transport, and return of the land with its remaining crops. These rules would be developed alongside a review of the institutional conditions for entrusted farming, actual support eligibility and agreements among participants.
Data management would also need to grow with participation. I would extend central administration, center-specific access and editing permissions, privacy protection and change histories to suit a larger network. Participating parcels' plans would be connected to, but kept distinct from, official national production, consumption and supply-and-demand data. Records accumulated across seasons could then help participating regions review crop concentration and resource needs together.
Once one area has built operating experience and farming results, I would explore whether that experience can transfer to other regions. The current software can check configured variety, trial-area and orchard conditions, but establishing a regional specialty still requires agronomic evidence, local cooperation and field results. Over time, the network could support suitable regional specialties and shared equipment use. Much further ahead, with owners' consent and the necessary institutional review, it could also provide a basis for discussing coordinated management and land consolidation among participating parcels.
Success would be assessed through more than farm income and working hours: whether land entrusted by older owners remains cultivated, essential work happens within the right seasonal windows, and local people can continue participating in agricultural work. The next goal is to establish an operating model that can endure in one area, then use that experience to develop a policy proposal for the long-term viability of Korean agriculture.
Built With
- nebius-token-factory
- next.js
- nvidia-nemotron
- or-tools
- postgresql
- prisma
- python
- react
- typescript
Log in or sign up for Devpost to join the conversation.