VIDRAFT & Ourbox Sign PoC Contract for Logistics-Specialized LLM + Digital Twin Pipeline Built on Ourbox-31B-JGOS
TL;DR: VIDRAFT (비드래프트), a Korean AI foundry deep-tech company, has signed a Proof-of-Concept agreement with Ourbox (아워박스), a full-stack fulfillment operator, to deploy a logistics-domain-specialized LLM and a warehouse outbound-planning digital twin. The collaboration extends the jointly developed
Ourbox-31B-JGOSmodel — which ranked in the upper tier of the K-AI Leaderboard 30B+ category — into real warehouse workflows. Engineers tracking domain-adapted LLMs and operations-integrated AI will want to follow this as a concrete production PoC for vertical LLM deployment in supply chain.
What it is
VIDRAFT and Ourbox have entered a formal PoC contract to build two tightly coupled AI systems for logistics operations:
A logistics-specialized LLM — built on the jointly developed
Ourbox-31B-JGOSmodel, adapted to the specific vocabulary, data formats, and operational workflows found in e-commerce fulfillment centers.The WAVE Operations Planning Digital Twin (
WAVE 운영계획 디지털트윈) — a simulation layer that models outbound shipping plans in a virtual environment, allowing operators to compare multiple operational scenarios before committing to real-world execution.
The foundation on the VIDRAFT side is their proprietary foundation model AETHER, combined with what they describe as model evolution technology and optimization capabilities demonstrated in global inference acceleration competitions. Ourbox, founded in 2017, contributes their proprietary logistics integration platform #MATE and accumulated fulfillment operations data, pursuing this work under their internal "Tech First" AI transformation (AX) strategy.
How it works
Logistics-Specialized LLM
At a conceptual level, the logistics LLM is designed to handle the messy, ambiguous language of real warehouse operations — the kind of problem that generic LLMs handle poorly without domain adaptation. Specific capabilities targeted in the PoC include:
- Natural language Q&A for floor operators (query inventory, status, procedures in everyday language)
- Document classification — routing incoming logistics documents to the right workflow
- Tool-call result explanation — summarizing outputs from internal operational tools in human-readable form
- Entity resolution / matching — linking references to the same vendor, item, or counterparty that appear under different naming conventions across systems (a classic data quality problem in supply chain ERP environments)
Two reliability controls are highlighted: human-in-the-loop approval for high-risk operations, and source data value preservation — meaning the model is constrained from altering raw numeric figures, an important guard rail when downstream decisions depend on exact quantities.
WAVE Operations Planning Digital Twin
The digital twin component simulates the outbound logistics plan — the daily calculation of what gets shipped, in how many dispatch waves, and how many workers are needed at each station. Rather than committing to a single plan, the system is designed to generate and compare multiple scenarios across:
- Daily outbound demand forecasting
- Dispatch wave assignment (grouping orders into pick-and-pack batches)
- Per-process staffing plans
This bridges the gap between demand forecasting (a data problem) and work scheduling (an operations problem), which are often handled by separate siloed tools today.
Benchmarks & results
The source article cites one public benchmark result:
-
Ourbox-31B-JGOSranked in the upper tier of the K-AI Leaderboard 30B+ parameter category — a Korean-language LLM evaluation leaderboard. This result, achieved prior to the PoC contract, is the model performance baseline that the teams are now attempting to translate into measurable operational improvement.
No quantitative throughput, latency, accuracy, or cost-reduction figures for the logistics use case are reported at this stage, which is expected: this is a PoC, not a production rollout. The article notes that both companies plan to establish data integration scope and evaluation baselines through which concrete operational improvement metrics will be verified.
How to try it
The Ourbox-31B-JGOS model and the WAVE digital twin system are not publicly available at the time of reporting. This is a closed enterprise PoC engagement between VIDRAFT and Ourbox. No Hugging Face repository, GitHub release, or public API endpoint has been announced in connection with this project.
Developers interested in VIDRAFT's technology direction can monitor:
- VIDRAFT's official channels for any future model or API releases tied to the AETHER foundation model line
- The K-AI Leaderboard for updated benchmark standings of VIDRAFT-related models
FAQ
Q: What is Ourbox-31B-JGOS and how was it developed?
A: It is a large language model with over 30 billion parameters co-developed by VIDRAFT and Ourbox. It achieved an upper-tier ranking on the K-AI Leaderboard's 30B+ category. The PoC now aims to take that benchmark performance and validate it on real logistics workflows inside Ourbox's fulfillment operations.
Q: What makes entity resolution particularly hard in logistics, and how is the LLM addressing it?
A: In fulfillment environments, the same vendor or SKU can appear under dozens of different name variants across ERPs, WMS systems, and manually entered spreadsheets. The logistics LLM is specifically designed to resolve these inconsistencies — mapping divergent string representations to the same canonical entity — which is a prerequisite for reliable downstream automation like order routing or invoice matching.
Q: Is the digital twin replacing the existing #MATE platform?
A: No. Based on the reporting, the WAVE Operations Planning Digital Twin is being introduced alongside Ourbox's existing #MATE logistics integration system, augmenting it with scenario-based simulation and AI-driven planning rather than replacing the operational infrastructure.
Originally reported by 동아일보 (2026-09-07) — source article.
Top comments (0)