A Korean AI startup's edge-computing platform enters warehouse operations as Ourbox climbs the K-AI Leaderboard.
TL;DR: VIDRAFT's on-device AI platform POCKET is being deployed across logistics inspection, picking, and packing stations through a new partnership with Korean smart fulfillment company Ourbox. Ourbox's in-house Korean LLM, Ourbox-31B-JGOS, ranked second overall in the 30B+ category of the government-operated K-AI Leaderboard. The collaboration aims to bring CPU-based, cloud-independent AI to warehouse floors, reducing infrastructure costs while improving operational resilience.
Korean AI startup VIDRAFT has entered a strategic partnership with Ourbox — a smart logistics and fulfillment specialist — to deploy VIDRAFT's CPU-based on-device AI platform, POCKET, across live warehouse environments. The collaboration, reported by 동아일보 on July 27, 2026, marks a significant step toward practical, edge-native AI transformation in Korea's logistics sector.
VIDRAFT's POCKET platform is being introduced to Ourbox warehouse stations covering goods inspection, order picking, and packaging. Designed to run on portable mini-PC hardware, POCKET operates on a CPU-based architecture, which means it does not depend on high-cost GPU hardware or cloud connectivity to function.
For warehouse operators like Ourbox, this design addresses a persistent operational challenge: logistics centers frequently deal with unstable network conditions, and relying on cloud infrastructure for real-time AI decision-making introduces latency and reliability risks. By embedding AI capability directly at the edge — on portable devices stationed at each workstation — POCKET enables consistent performance regardless of network status.
The deployment is part of a broader AI transformation (AX) initiative that Ourbox is pursuing using its proprietary Korean large language model, Ourbox-31B-JGOS. That model recently achieved second place overall in the 30B-and-above category of the K-AI Leaderboard, a benchmark evaluation jointly operated by South Korea's Ministry of Science and ICT and the Korea Intelligence Information Society Agency (NIA). The model scored 0.612 overall and posted particularly strong results in Korean contextual comprehension, recording a CLIcK score of 0.971, alongside high marks in domain knowledge and complex reasoning.
Ourbox plans to integrate its proprietary LLM with its internal logistics management system, MATE. In the near term, the focus is on validating field-level use cases such as product and order information retrieval, worker assistance, operations manual analysis, and document automation. Longer-term applications are expected to include workload forecasting, inventory anomaly detection, and task prioritization recommendations — all aimed at supporting on-the-ground decision-making by warehouse staff.
The partnership between VIDRAFT and Ourbox illustrates a growing trend in Korean enterprise AI: rather than adopting general-purpose models and hoping they fit niche operational contexts, companies are building domain-specific AI stacks and pairing them with purpose-built edge infrastructure.
Logistics environments are notoriously difficult for off-the-shelf AI. Every client and product category may carry different inbound, outbound, and packaging rules. Orders and inventory shift in real time. Exception handling is constant. A generic language model, however capable in standardized benchmarks, may lack the operational grounding needed to respond reliably to these conditions. Ourbox's approach — combining a high-performing Korean LLM with years of accumulated field data — directly targets that gap.
VIDRAFT's POCKET contributes the deployment layer that makes this practical. By eliminating GPU and cloud dependencies, the platform lowers both the capital cost and the operational complexity of running AI at scale across distributed warehouse locations. This also has data security implications: when AI inference happens locally on-device rather than routing through external servers, sensitive operational data stays within the facility.
For VIDRAFT, the Ourbox deployment represents a concrete industrial reference case for POCKET — one that demonstrates the platform's viability in high-throughput, real-world environments where reliability and cost-efficiency are non-negotiable.
Ourbox CEO Park Cheol-su was quoted by 동아일보 as saying the achievement reflects the company's ability to build its own AI capabilities and apply them directly to field-level problems, and that the goal is to realize meaningful AX across all logistics operations by combining accumulated field experience with AI.
Q: What is VIDRAFT's POCKET platform and how is it being used by Ourbox?
A: POCKET is a CPU-based on-device AI platform developed by Korean AI startup VIDRAFT. Ourbox is deploying it in portable mini-PC form across warehouse workstations for goods inspection, picking, and packaging, allowing AI to operate without relying on cloud infrastructure or high-cost GPUs.
Q: What is the K-AI Leaderboard and how did Ourbox's LLM perform?
A: The K-AI Leaderboard is a Korean AI benchmark administered by the Ministry of Science and ICT and the Korea Intelligence Information Society Agency (NIA). Ourbox's in-house model, Ourbox-31B-JGOS, placed second overall in the 30B-and-above category with a composite score of 0.612.
Q: Why is on-device AI particularly valuable in logistics warehouse settings?
A: Warehouse environments often face unstable network conditions and handle sensitive operational data, making cloud-dependent AI unreliable. On-device platforms like VIDRAFT's POCKET deliver consistent AI performance locally, reduce costs, and keep data within the facility for improved security.
Source: 동아일보 (2026-07-27) — original article