Two Korean tech companies are moving AI out of the lab and onto the warehouse floor.
TL;DR: VIDRAFT, a Korean AI startup, has signed a proof-of-concept agreement with fulfillment company Ourbox to deploy a logistics-specialized large language model and an outbound-planning digital twin at live warehouse operations. The collaboration builds on the jointly developed Ourbox-31B-JGOS model, which ranked among the top entries in the 30B-and-above category of the K-AI Leaderboard. Both companies announced the deal on September 7, 2026.
Korean AI startup VIDRAFT and comprehensive fulfillment provider Ourbox formally announced on September 7, 2026, that they have signed a proof-of-concept (PoC) contract to bring logistics-specialized artificial intelligence into real warehouse environments. The agreement covers two integrated systems: a domain-specific large language model (LLM) tailored to logistics operations, and a simulation-based digital twin called the "WAVE Operations Planning Digital Twin."
The PoC contract marks the operational expansion of work the two companies already started together. VIDRAFT and Ourbox previously co-developed a model named Ourbox-31B-JGOS, which achieved a top-tier ranking in the 30B-and-above segment of the K-AI Leaderboard. The new agreement takes that research-grade achievement and translates it into practical, day-to-day logistics functionality.
Logistics-specialized LLM
The domain-specific LLM is designed to handle the kinds of tasks that warehouse staff encounter constantly. These include natural-language Q&A in everyday operational language, document classification, explanation of results returned by work-tool calls, and matching of item and counterparty information. A particularly notable capability is the model's ability to recognize that different text inputs — different spellings, abbreviations, or internal codes — refer to the same trading partner, linking them automatically. The system also applies human-in-the-loop controls for high-risk actions, requiring supervisor approval before execution and preserving source data figures to maintain auditability and trust.
VIDRAFT brings to this effort its proprietary foundation model, AETHER, along with model-evolution techniques and optimization capabilities it has demonstrated in global inference-acceleration competitions. The company describes itself as an AI foundry and deep-tech firm.
WAVE Operations Planning Digital Twin
The second component, the WAVE digital twin, creates a virtual replica of outbound logistics flows at a fulfillment center. Planners can use it to simulate daily shipment volume forecasts, assign outbound waves, and draft staffing plans across each stage of the operation — all within a multi-scenario comparison framework. The goal is to close the gap between demand forecasting and executable work plans, giving operations managers a structured way to test decisions before committing resources.
Ourbox, founded in 2017, operates its own integrated logistics management system called #MATE and handles end-to-end e-commerce fulfillment. The company has recently been pursuing an AI transformation (AX) strategy it calls "Tech First," and the VIDRAFT partnership is a central piece of that push.
Logistics is widely recognized as a sector where AI can deliver measurable value, yet it is also one of the hardest domains to crack. Operations vocabulary is highly specialized, data is often messy and inconsistently formatted, and errors carry real physical consequences — a mis-picked order, a miscounted pallet, an understaffed shift. That combination makes it an acid test for whether a specialized LLM can outperform a general-purpose one.
VIDRAFT CEO Kim Min-sik addressed this directly, noting that logistics — with its distinctive terminology and data — is precisely the kind of environment where the value of a domain-specific model gets proven. Ourbox CEO Park Cheol-su echoed that framing, stating that the company intends to combine its accumulated logistics operations experience and data with AI in order to advance how it runs its business.
The PoC structure is deliberately staged. Both companies say they will define data-integration scope and evaluation baselines first, then proceed through a formal verification and integration process before drawing conclusions about operational improvements. That measured approach reflects growing industry awareness that splashy AI announcements need to be followed by rigorous, criteria-based validation.
For the broader Korean AI ecosystem, the deal is a signal that domestic foundation-model builders like VIDRAFT are moving beyond benchmark competitions and into production deployments — a transition that will ultimately matter far more for the industry's long-term credibility.
Q: What is the Ourbox-31B-JGOS model?
A: Ourbox-31B-JGOS is a large language model co-developed by VIDRAFT and Ourbox that ranked in the top tier of the 30B-and-above category on the K-AI Leaderboard. The new PoC agreement is designed to translate that benchmark performance into practical logistics applications.
Q: What will the logistics-specialized LLM actually do in the warehouse?
A: The model will handle natural-language queries from warehouse staff, classify documents, explain the outputs of operational tools, and match item and counterparty data — including recognizing the same trading partner referenced under different names or codes. High-risk actions require human approval before the system proceeds.
Q: What is the WAVE Operations Planning Digital Twin?
A: It is a simulation environment that mirrors a fulfillment center's outbound logistics flows. Planners can use it to compare multiple scenarios for daily shipment volume forecasts, outbound wave assignments, and per-process staffing levels, helping connect demand predictions to concrete work plans before they are put into practice.
Source: 동아일보 (2026-09-07) — original article