How SF Tech’s Proprietary Large Models Revolutionize Logistics and AI Operations

Recently, the DataFun community hosted the DA Data Intelligence Conference in Shenzhen, focusing on big data and AI applications. Experts from Amazon Cloud, Xiaohongshu, Alibaba Cloud and others presented on large‑model usage and data architecture evolution across logistics, insurance, social media, and software development.

SF Tech’s CMO Tang Kai delivered a keynote titled “Supply‑Chain Intelligent Decision‑Making with Fengyu and Fengzhi Large Models,” outlining a custom logistics‑focused large‑model pipeline. He noted the limitations of generic models for specialized domains and introduced the multimodal “Fengyu” model and the “Fengzhi” logistics decision model, covering customer, operations, international logistics, supply‑chain, and smart office scenarios.

The large‑model system is already deployed at scale in SF’s business, delivering benefits such as demand forecasting, decision optimization, and operational analysis. In sectors like beauty, 3C, food, and auto parts, the model reduced server resource demand by eightfold, increased runtime efficiency by 120×, and improved prediction accuracy by 5%.

Multimodal models also boost operational optimization, international business, and customer experience. One‑sentence ordering, image‑based order creation, and rapid key‑information extraction cut average customer‑service handling time by 30%. In international operations, automated customs‑document review now handles over 97% of cases, markedly raising efficiency and service quality.

SF Tech’s Cognitive Decision‑Making Agent applies a self‑developed vertical large model combined with spatio‑temporal network prediction, large‑scale optimization, and dynamic capacity scheduling. It supports logistics network planning, demand assessment, and capacity‑resource improvement, offering an explainable, high‑accuracy decision framework.

In aviation, the agent can resolve abnormal scheduling through natural‑language intent parsing, scenario matching, algorithm invocation, and result interpretation, dramatically improving response speed.

RAG Innovation by Senior Algorithm Engineer Lu Yue addresses hallucination issues with three breakthroughs: a real‑time dynamic retrieval mechanism (DRAGIN framework with RIND and QFS modules) that raises security‑alert handling accuracy to 96% and cuts false‑positive rates by 14%; controllable generation via RETRO‑based chunked cross‑attention and GCR‑based knowledge‑graph constraints achieving 99.3% audit consistency; and an end‑to‑end joint training paradigm (ATLAS framework) that lifts sensitive‑data classification accuracy to 95%.

These RAG techniques are being applied in SF’s information‑security domain to ensure reliable LLM deployment.

Backend Engineer Song Chengchuan presented the enterprise‑grade Agent Platform. Built on a custom eGPU pool, hybrid‑cloud optimization, and resource scheduling, the platform (based on a heavily customized Dify) integrates an internal model marketplace, delivering over 10 billion daily token calls with unified authentication, load balancing, security auditing, and protocol conversion.

The platform includes three key Agentic tools: a Model Marketplace for private and commercial model services; an Evaluation Platform offering performance and capability benchmarks (latency, resource consumption, accuracy, multimodal ability) via offline/online and rule‑based or judge‑model scoring; and an Observability Platform (leveraging Langfuse) for monitoring, root‑cause analysis, and SLA assurance.

Product Operations Manager Lu Xinting shared the end‑to‑end lifecycle of the “Fengyu” model, emphasizing a product‑centric approach: model verification, content polishing, data‑driven iteration, and ecosystem co‑creation, while tackling hallucination mitigation and cost control. The “Xiao Ge Service Center” chatbot, powered by Fengyu, has handled over 6 million conversations with a 90.41% problem‑resolution rate and a 20% increase in product penetration within six months.

Overall, SF Tech continues to drive AI‑driven logistics transformation, leveraging large‑model technology as a core engine for cost reduction, efficiency gains, and revenue growth across the supply‑chain.