Eurocontrol Uses DesignVerse AI to Modernize European Air Traffic Stack

欧洲空管组织采用DesignVerse AI推进空中交通技术栈现代化 该欧洲空中交通管理机构利用领域专用AI在生产环境中直接重写传统控制系统，在避免运营停机风险的同时将软件交付成本降低50%。 EUROCONTROL每年管理约1000万架次欧洲航班，但仍依赖于一套拥有30年历史的传统软件技术栈，任何非计划停机都会带来巨大的财务和运营成本。为化解这一技术债务，EUROCONTROL与初创公司DesignVerse合作，部署了连接用户界面设计系统、航空领域逻辑和后端代码生成的专用上下文层。在为期1.5个月的试点项目中，双方将一个正在运行的空域地图应用程序现代化升级至React和Java架构；与通用AI编程工具相比，工程代码产出翻倍，交付成本降低了50%。展望未来，该举措旨在通过自然语言工具连接业务专家与开发人员，并通过部署在本地的参数量低于1000亿的小型领域专用LLM来确保数字主权。 欧洲航空安全组织（EUROCONTROL）管理着覆盖欧洲多个主权国家的空域网络，每年协调超过1000万架次航班，单次非计划延误每分钟造成约100欧元损失。 (0:04) 针对拥有30多年历史的专有老旧系统INM，EUROCONTROL与初创公司DesignVerse合作，通过构建操作上下文层来规避通用AI编程工具在强监管领域的局限。 (0:50) 在空域地图应用的试点项目中，团队仅用1.5个月完成向React和Java架构的全栈迁移，使交付成本降低50%，开发产出提升至2倍。 (8:44) DesignVerse平台支持非技术人员通过自然语言实时同步修改前端UI和后端逻辑，生成80%至90%达到生产就绪级别的代码。 (12:03) EUROCONTROL计划利用自然语言接口打通领域专家与软件工程师，而DesignVerse则致力于向银行与网络安全等强监管领域推广该现代化框架，并研发本地部署的小型专用大模型保障数据主权。 (17:24)

To modernize Europe's air traffic network without disrupting live operations, EUROCONTROL is combining general-purpose language models with domain-specific AI tooling to refactor decades of legacy software while protecting operational continuity and data sovereignty. High Operational Stakes and Legacy Technical Debt (0:04) EUROCONTROL coordinates airspace across multiple European sovereign states, managing roughly 10 million flights each year with peak volumes reaching 37,000 daily flights. According to Innovation Program Manager Razvan Marginean, the agency operates under extreme operational constraints: every single minute of system delay creates approximately €100 ($115 USD) in costs passed down to airlines and travelers. However, the organization relies on a single, in-house software stack accumulated over more than 30 years. Through its Integrated Network Manager (INM) initiative, EUROCONTROL is attempting to migrate these critical workflows to cloud-native systems without incurring unscheduled downtime. The agency identified two primary engineering bottlenecks: long requirement-gathering cycles between operational domain specialists and developers, and the slow, risky process of converting legacy codebases into modern software languages. Pilot Modernization and Engineering Outcomes (8:44) Rather than deploying AI directly to mission-critical flight-control cores immediately, EUROCONTROL targeted engineering practices and auxiliary operational tools. The agency partnered with DesignVerse, an AI software startup founded in 2024, to modernize an operational information board used by air traffic controllers to track sector regulations and airspace overlaps. The existing tool—built in Rust, HTML, and CSS—was rewritten into a modern React and Java architecture aligned with EUROCONTROL's standardized design system. According to DesignVerse CEO Andrei Manolache, the front-end and back-end modernization finished in 1.5 months. Manolache reported a 50% drop in delivery costs and a twofold increase in code output compared to standard tools like Claude Code or Codex. He noted that the platform generates full-stack code that is roughly 80% to 90% production-ready, allowing product managers and architects to revise application logic simultaneously through natural language prompts. Hybrid AI Strategy and Operational Alignment (13:16) EUROCONTROL established an internal AI Center of Excellence relying on a two-tier hybrid architecture. General-purpose large language models, such as Claude, assist the broader 5,000-employee workforce with routine software tasks, while specialized tools provide the deeper context, design rules, and architectural constraints necessary for aviation compliance. Marginean emphasized that this visual, prompt-driven interface helps close the communication divide between non-technical airspace specialists and software engineers. Over a ten-year horizon, EUROCONTROL envisions fully harmonizing its stakeholder-facing operational product suite through unified natural language interfaces. Digital Sovereignty and On-Premise Models (19:21) Moderator Nancy Scola raised concerns about European organizations relying on foreign, US-based commercial models that may be vulnerable to regulatory or export restrictions. In response, Manolache noted that DesignVerse is developing smaller, domain-specific models under 100 billion parameters designed for local, on-premise execution. Manolache claimed that these specialized architectures, when provided with rich contextual frameworks, can match or exceed the performance of massive general-purpose models on focused tasks. Running sub-100-billion-parameter models on commodity, consumer-grade GPU hardware allows organizations in sensitive sectors like aviation, banking, and cybersecurity to retain full intellectual property and data sovereignty. While EUROCONTROL achieved significant speedups in its pilot, broader transformation across its 30-year legacy footprint remains an ongoing, multi-year process. The agency's long-term vision depends on scaling domain-specific context layers without compromising regulatory compliance or system uptime.

欧洲空中航行安全组织（EUROCONTROL）正通过与初创公司 DesignVerse 合作，利用具备深层领域上下文的专用 AI 工具对其积攒逾 30 年的遗留空管软件进行现代化改造。该方案在保障实时系统不中断并降低巨额延误成本的前提下，试图克服通用大型语言模型在强监管环境中的落地局限，同时探索依托本地化小型模型的数字主权路径。 遗留架构债务与高昂延误成本 (0:50) EUROCONTROL 负责协调欧洲多国空域网络，每年统筹处理约 1000 万架次航班，高峰期单日达 3.7 万架次。该机构长期依赖一套由内部自主维护超过 30 年的高度定制化单一软件栈，其中充斥着专有技术与利基编程语言。其启动的“集成网络管理器”（INM）计划旨在平稳迁移至云原生架构，但面临极高的技术与运维风险，因为核心调度系统绝不允许出现非计划停机。 EUROCONTROL 创新项目经理 Razvan Marginean 指出，系统性能与运营成本直接挂钩，空管系统每产生 1 分钟延误，就会给航司及旅客带来约 100 欧元（115 美元）的直接损失。为此，该组织未在核心飞行系统上贸然进行重构，而是优先将 AI 引入工程规范与交付流程，旨在缩短非技术业务人员与工程团队间的需求转化周期，加速淘汰过时的专用开发工具。 空域地图试点与全栈代码生成 (8:44) 在首个现代化试点中，团队针对空管人员用于监测欧洲空域、管制扇区及航线重叠的操作信息板展开重构。该项目将基于旧版 Rust、HTML 和 CSS 的应用重写为现代化的 React 与 Java 架构，并深度适配专为航空领域打造的 EUROCONTROL 官方设计系统。DesignVerse 首席执行官 Andrei Manolache 强调，Cursor、Claude Code 等通用编程工具难以处理航空业严格的领域约束，而专用工具能通过领域上下文层实现界面与后端业务逻辑的同步生成。 项目实测数据显示，该空域地图的端到端现代化改造仅耗时 1.5 个月，不仅将交付成本降低了 50%，代码产出效率更达到通用编码工具的两倍。这一结论直接基于布鲁塞尔生产服务器的真实代码库评测，而非通用的公开基准测试（如 SWE-bench）。Manolache 称，系统生成的代码生产就绪度达到 80% 至 90%，能够直接对接开发团队现有的 CI/CD 持续集成流水线。 自然语言协作与混合 AI 布局 (13:16) 航空软件开发面临的一大瓶颈是空域规则专家与代码编写人员之间的认知壁垒。EUROCONTROL 借助实时可视化反馈与自然语言提示交互，允许产品经理、架构师及一线管制专家直接调整并验证全栈逻辑，从而大幅缩短了前期的需求梳理周期。Marginean 透露，其长达 10 年的战略目标是通过简化的统一工具和自然语言接口，彻底整合面向各类利益相关方的全部操作产品线。 为了兼顾规范严谨性与整体创新效率，EUROCONTROL 设立了内部 AI 卓越中心并推行混合技术路线。该机构在其拥有 5000 名员工的组织内部部署了 Claude 等通用大型语言模型，用以增强常规研发效能；而在涉及严格行业标准、复杂架构映射以及合规要求的专业场景中，则采用由 DesignVerse 提供的持久化航空领域上下文模型进行对齐。 数字主权防范与本地化专有模型 (19:21) 针对外部商业大模型可能受制于外国出口管制或地缘政治政策变动的风险，DesignVerse 主张客户必须对自身运营数据和数十年的智力资产保有绝对控制权。为了落实真正的数字主权，该公司正致力于开发参数量低于 1000 亿（<100B）的领域专用模型，支持完全脱离公有云、在客户本地环境独立运行。 Manolache 认为，在精确的业务上下文和架构规范支撑下，这类轻量化专有模型在特定企业任务中的代码生成质量与效率可超越通用超大规模模型（如 GPT、Claude 或 Qwen），并且仅需依托易于批量采购的消费级或通用 GPU 即可平稳运行，满足极高安全标准。除航空管制领域外，该机构与团队亦计划将该架构向国家级空中航行服务提供商（ANSP）、银行及网络安全等受制于遗留系统的其他高门槛行业推广。 该试点验证了通过领域上下文增强与自然语言全栈生成来加速工业遗留系统重构的可行性。然而，该方案能否从边缘信息板顺利扩展至容错率更低的核心飞行调度体系，以及消费级硬件承载的本地化小模型在更复杂的监管场景中能否稳定替代顶级商用大模型，仍有待更大规模的工程实证。

## Background 本内容由 YouTube 频道 Ai4 于 2026 年 8 月 15 日发布。视频围绕欧洲空中交通管理机构 EUROCONTROL 如何在不中断 24/7 实时运行的前提下，利用 AI 初创平台 DesignVerse 对其积累 30 年的技术债与遗留系统进行现代化改造展开深入探讨。 ## TL;DR EUROCONTROL 面临着每年协调 1000 万架次航班且每延误一分钟损失约 100 欧元的巨大运营压力，其 30 年前构建的复杂遗留系统亟需改造。通过与初创公司 DesignVerse 合作，团队利用注入航空专有上下文与设计系统的 AI 平台，仅耗时 1.5 个月便完成了信息看板系统从 Rust + HTML/CSS 向 React + Java 的端到端重构，交付成本降低 50% 且工程效率翻倍。该方案不仅通过自然语言全栈修改弥合了业务专家与开发人员之间的鸿沟，还探索了参数规模小于 100B 的私有化轻量模型路线，以保障数据安全与数字主权。 ## Atmosphere & style 对话整体呈现出务实、严谨且充满前瞻性的专业科技访谈氛围。资深科技记者 Nancy Scola 的提问直击受监管行业转型的痛点与地缘政策考量，而两位嘉宾的对谈则在深厚的工程现实（如 30 年技术债、严苛的合规与不关机要求）与创新的 AI 落地实践之间找到了极佳的平衡，既有严密的技术逻辑推演，又展现了高风险工业系统拥抱前沿 AI 时的沉稳与笃定。 ## Key takeaways - [00:59](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=59s) EUROCONTROL 负责协调欧洲全空域的航班流量与容量，高峰日处理高达 37,000 架次航班，每延误一分钟的系统成本约为 100 欧元。 - [00:59](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=59s) EUROCONTROL 的核心系统积累了超过 30 年的定制化技术债务，且必须保持 24/7 全天候在线运行，无法停机改造。 - [00:59](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=59s) 启动 INM（综合网络管理器）数字化转型项目，核心在于利用 AI 加速需求转化并将老旧安全架构平稳迁移至现代化云原生技术栈。 - [06:13](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=373s) EUROCONTROL 采取渐进式风险控制策略，先从周边运营支撑看板系统验证 AI 现代化工程流程，再逐步向核心飞行管理系统推进。 - [07:46](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=466s) DesignVerse 仅耗时 1.5 个月即完成首个信息看板系统从 Rust + HTML/CSS 到 React + Java 的端到端重构，复用了 EUROCONTROL 自研航空专用设计系统。 - [07:46](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=466s) 相比通用辅助编码工具，该垂直 AI 现代化引擎将交付成本降低 50%，工程产出效率提升 2 倍，并具备组织级代码习惯的复合学习效应。 - [11:02](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=662s) 平台支持通过自然语言 Prompt 同步修改部署在生产服务器上的全栈代码，直接交付完成度达 80% 至 90% 的可用应用，大幅弥合业务专家与研发团队的鸿沟。 - [14:29](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=869s) EUROCONTROL 设立 AI 卓越中心推进全局治理，强调“模型中立与领域上下文为王”，通过专有上下文层承载深奥的空管业务约束。 - [17:32](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=1052s) 这套垂直现代化方案不仅服务于全球航空管制机构（ANSP），未来还将横向推广至银行金融、网络安全等强监管与高遗留债务行业。 - [18:24](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=1104s) EUROCONTROL 的十年愿景是通过自然语言交互与统一现代化层打通运营与技术认知，构建一体化、标准化的协调产品栈。 - [19:21](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=1161s) 为保障数字主权与合规，DesignVerse 研发参数在 1000 亿（100B）以内的轻量化专有模型，支持在消费级 GPU 上完全私有化部署并摆脱外部供应链依赖。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=4s) 航空业的高风险数字化转型：EUROCONTROL 与 DesignVerse 的合作背景 来自华盛顿特区的资深科技记者 Nancy Scola 为本次讨论拉开帷幕。她指出，欧洲空中交通管制网络 EUROCONTROL 正在通过 AI 实现系统现代化的案例极具启发性。航空业属于典型的高风险、容错率极低的行业，如何在保证运营安全的前提下顺利部署 AI 是整个行业的核心关切。 本次对话的两位嘉宾分别是：2024 年成立于布加勒斯特、专注于受监管行业 AI 软件开发的初创公司 DesignVerse 首席执行官 Andrei Manolache，以及 EUROCONTROL 的创新项目经理 Razvan Marginean。 - **EUROCONTROL**：负责欧洲全空域空中交通流量与容量协调的网络管理机构。 - **DesignVerse**：专注于将软件架构、设计系统与运行时逻辑相连接的 AI 现代化平台。 --- ### [00:59](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=59s) 30 年技术债与永不关机的系统：EUROCONTROL 的现代化难题 Razvan Marginean 介绍了 EUROCONTROL 承担的庞大网络管理职责。与具体的空中交通管制员（ATC）不同，EUROCONTROL 负责协调跨越欧洲各成员国、航空公司及管制中心的整个空域网络，每年有约 1000 万架次航班穿过其空域，在高峰日每天处理多达 37,000 架次航班。这一机制类似于美国的 FAA，需要统筹跨国界的大量利益相关方，形成统一运作的空域。 网络管理最核心的运营目标之一是尽可能压缩延误时间，因为在航空业中“每延误一分钟的成本大约是 100 欧元（约合 115 美元）”，这些成本最终会层层传递给航空公司及乘客。EUROCONTROL 支撑业务的核心软件系统积累了超过 30 年的工程技术债务，是一套高度定制化的专用系统，而非通用的商用现成软件。最大的挑战在于系统必须保持 24/7 全天候在线运行，“在进行现代化改造的同时，我们绝对不能将系统关闭”。 ```mermaid flowchart TD accTitle: EUROCONTROL INM Digital Transformation Flow subgraph OperationalContext["运营与监管层"] A["多国监管与利益相关方"] --> B["航班流量与空域协调"] B --> C["延误成本控制: ~100欧元/分钟"] end subgraph Modernization["INM 综合网络管理改造"] D["30年专有遗留架构"] --> E["需求到代码加速 (AI协作)"] D --> F["上云与新一代技术栈迁移"] end OperationalContext --> Modernization ``` 为了应对这一挑战，EUROCONTROL 启动了名为 INM（Integrated Network Manager，综合网络管理器）的数字化转型项目。该项目聚焦于两大 AI 应用方向：一是通过 AI 加速跨部门、多利益相关方的需求梳理与转化周期；二是平稳将古老但经过安全验证的旧技术栈迁移至现代化架构。 --- ### [06:13](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=373s) 破局之道：为何选择一家仅成立两年的 AI 初创企业？ 当主持人 Nancy Scola 问及为何愿意将极其关键的系统现代化改造委托给成立仅两年的初创团队时，Razvan Marginean 解释了 EUROCONTROL 的考量。数字化转型推进到攻坚阶段时，团队亟需显著提升交付效能与开发敏捷度。市场现存的通用企业软件很难契合航空业这种极其垂直、复杂的场景，而 DesignVerse 的理念精准击中了“压缩从业务需求到生产上线全周期”这一核心痛点。 EUROCONTROL 采取了循序渐进的风险控制策略：并没有一上来就动用最核心、最敏感的“皇冠上的明珠”（如涉及严格适航认证的核心飞行管理系统），而是先从周边运营支撑系统入手，探索并验证现代化的工程方法与治理流程，在流程和规范稳固后，再逐步向核心业务系统推广。 --- ### [07:46](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=466s) 走出实验室：从 Rust + HTML/CSS 到 React + Java 的实战重构 Andrei Manolache 回忆起刚接手项目时的感受：“任何人都明白，航空、金融、医疗这类受监管行业中，充斥着服役 10 年、20 年乃至 40、50 年的古老软件。”通用型 AI 辅助编码工具（如 Claude Code、Windsurf、Codex、Cursor 等）在小型企业或通用场景下表现良好，但在极其垂直的硬核工业体系中却力不从心。DesignVerse 的目标正是证明 AI 可以在最严苛的航空管制场景中扎根。 首个落地项目是面向管制员的信息看板系统（用于查看全欧洲空域状态、机场起降时刻 Slot 及流控限制等）。改造目标是将原有的 Rust + HTML/CSS 架构重构为 React + Java。前端选用 React 是为了直接复用 EUROCONTROL 投入大量资源研发的自研航空专用设计系统（Eurocontrol Design System）。 ```mermaid flowchart LR accTitle: Migration Pipeline Architecture subgraph LegacyStack["遗留系统"] L1["Rust 后端"] L2["传统 HTML/CSS 前端"] end subgraph DesignVersePlatform["DesignVerse AI 现代化引擎"] DV1["自然语言端到端改写"] DV2["Eurocontrol 航空专用设计系统"] end subgraph ModernStack["现代化生产栈"] M1["Java 企业级后端"] M2["React 前端组件"] end LegacyStack --> DesignVersePlatform DesignVersePlatform --> ModernStack ``` - **重构周期**：仅耗时 1.5 个月即完成了前端与后端的端到端重构。 - **降本增效**：与基准工具（如 Claude Code 等）相比，交付成本降低了 50%，代码产出工程效率提升了 2 倍。 - **复利效应**：企业级基准测试完全基于真实业务代码库，随着系统不断理解 EUROCONTROL 5000 多名员工背后的架构规范与编写模式，AI 产生了持续进化的复合学习效应。 --- ### [11:02](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=662s) 双向打通：自然语言全栈修改与业务/技术鸿沟的弥合 Andrei 展示了系统改造前后的对比：旧界面风格老旧、缺乏连贯的 UX 交互；现代化重构后的应用则清晰展示了欧洲各扇区（Sectors）、行政辖区（如瑞士的各州）的空域管制与冲突排查信息。 DesignVerse 的核心突破在于提供了一个面向产品与验证的实时运行时平台。用户在 Web 界面中通过自然语言发出 Prompt，系统会同步修改部署在布鲁塞尔服务器上的真实前后端代码。这与市面上仅生成前端玩具原型的工具有着本质区别：“经过用户验证后，交付给开发团队的是一套完成度高达 80% 至 90% 的端到端可用应用”，开发人员只需在此基础上补充 QA 测试、接入 CI/CD 并通过预发环境即可上线。 Razvan 强调，该工具最大的价值在于架起了“懂运营不懂代码”的业务专家与“懂代码不懂航管”的技术工程师之间的桥梁。可视化的实时产品层让双方能够在一致的上下文中高效协作，大幅消除了沟通损耗与返工风险。 --- ### [14:29](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=869s) 深度协同与全景式 AI 治理战略 针对如何确保系统精确复刻复杂的空管知识体系，Andrei 指出，团队在早期与 EUROCONTROL 紧密配合，深入底层了解其 30 年积累的工程架构逻辑，并将这些业务规则沉淀为 Agent 与 AI Pipeline 的上下文基座，逐步实现更高阶的自动化。 面对 Nancy 关于为何不单纯依赖大型科技公司通用大模型的提问，Razvan Marginean 阐明了 EUROCONTROL 的整体 AI 布局： - **设立 AI 卓越中心（AI Center of Excellence）**：从全局视角评估与推进全组织的 AI 实践，避免押注单一工具或单一技术路径。 - **模型中立与领域上下文为王**：底层的代码生成可以使用任何通用 LLM，但通用大模型无法内生理解并持久保留欧洲空管深奥的业务约束，必须通过专有的上下文层进行强化。 - **双轨并行策略**：一方面利用通用工具提升常规软件开发效率，另一方面利用具备深厚行业上下文的平台（如 DesignVerse）承载业务逻辑与架构重塑，稳步推进上云与技术演进。 ### [17:32](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=1052s) 拓展行业版图与跨领域系统现代化 当被问及这套现代化解决方案是仅服务于全球其他空中交通管制系统，还是能横向拓展到航空之外的行业时，Andrei Manolache 表示这其实是两者的结合。DesignVerse 正依托在航空这一垂直业务支柱领域积累的深厚专识，推广给更多亟需系统现代化的客户。不仅将 EUROCONTROL 视为客户，双方更在紧密合作，共同向整个行业分享 EUROCONTROL 过去 20 年间沉淀的架构经验与系统构建知识，帮助监管机构以及负责各国本土交通管制的空中交通导航服务商（ANSP）实现系统升级。 从公司长期定位来看，DesignVerse 的目标是成为面向各类企业级应用场景的垂直现代化解决方案。除了当前深耕的航空领域外，银行业与网络安全亦是其核心支柱。未来，只要涉及复杂软件工程开发与遗留遗留系统（Legacy Applications）改造的行业，这套方法论与工具链都能灵活适配并落地。 - **三大核心垂直领域**：航空（Aviation）、银行金融（Banking）与网络安全（Cybersecurity）。 - **行业协同赋能**：将 EUROCONTROL 20 年的架构积淀标准化，向全球监管机构及各地 ANSP 推广。 --- ### [18:24](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=1104s) EUROCONTROL 的十年远景与一体化产品栈 针对项目从概念验证（PoC）走向长期落地的规划，Nancy Scola 询问了 EUROCONTROL 未来十年的宏大愿景。Razvan Margauan 坦言，在不透露过多核心战略的前提下，他最期待看到的是将面向各类运营客户与利益相关方的繁杂产品进行统一与协调（Harmonization）。通过引入更简便的工具乃至自然语言交互界面，进一步弥合深度技术知识（Technical know-how）与一线运营知识（Operational know-how）之间的鸿沟，打造一个真正一体化的产品栈（Integrated product stack）。 Razvan 强调，这是一场宏大的长期数字化转型工程，绝不可能在一夜之间一蹴而就。目前团队的核心任务是扎实搭建起底层的治理与技术框架，确保当系统进入大规模扩展阶段时，能够精准匹配高标准的业务与安全要求。 ```mermaid flowchart TD accTitle: EUROCONTROL 现代化演进路径 subgraph Gap["核心挑战"] A["一线运营知识<br/>(Operational Know-how)"] B["底层技术架构<br/>(Technical Know-how)"] end subgraph Solution["统一现代化层"] C["自然语言与简易交互工具"] D["底层标准治理框架"] end subgraph Goal["未来十年愿景"] E["协调一致的产品栈<br/>(Harmonized Product Stack)"] F["面向全利益相关方的敏捷运营支持"] end A & B --> C C --> D D --> E --> F ``` --- ### [19:21](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=1161s) 数字主权与百亿级本地轻量化模型 在访谈尾声，Nancy Scola 从科技政策与地缘政治视角切入，探讨了自建定制化模型与“数字主权”（Digital Sovereignty）的关系——即避免过度依赖可能受制于出口管制政策的美国主流模型提供商（如 Anthropic 等）。对此，Andrei Manolache 给出了清晰的技术路线回应。他指出，DesignVerse 的终极目标是赋能客户完全掌控自己的数据资产以及过去 10 到 20 年所积累的业务逻辑。 ```mermaid flowchart LR accTitle: 私有化轻量模型 vs 通用超大模型 subgraph Public["通用公有云大模型"] A["千亿/万亿级参数大模型<br/>(Claude / GPT 等)"] B["外部云端托管 / 存在出口管制风险"] end subgraph Private["DesignVerse 垂直专属方案"] C["< 100B 参数专属模型"] D["注入精准业务上下文与规则"] E["消费级 GPU 即可私有化部署<br/>(On-Premise)"] end A -. 依赖外部API.-> B C --> D --> E ``` 与其依赖庞大且黑盒的通用模型，DesignVerse 致力于研发可本地私有化部署（On-premise）的专有轻量化大语言模型： - **专注轻量化与高效率**：模型参数规模严格控制在 1000 亿（100B）以内。实践证明，在拥有精准上下文与规范指南的约束下，针对特定业务编写代码时，小型定制模型的效率往往远优于直接调用通用模型。 - **硬件门槛低与供应链可控**：百亿级模型可在消费级 GPU 上顺畅运行，硬件可从公开市场轻松批量采购，彻底摆脱单点供应链制约。 - **高敏感数据安全落地**：极度简便的本地部署能力，能够满足航空、金融等行业对敏感数据不出域的严苛合规要求。 ## Notable quotes > [00:59](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=59s) 在进行现代化改造的同时，我们绝对不能将系统关闭 — Razvan Marginean > [00:59](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=59s) 每延误一分钟的成本大约是 100 欧元（约合 115 美元） — Razvan Marginean > [07:46](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=466s) 任何人都明白，航空、金融、医疗这类受监管行业中，充斥着服役 10 年、20 年乃至 40、50 年的古老软件。 — Andrei Manolache > [11:02](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=662s) 经过用户验证后，交付给开发团队的是一套完成度高达 80% 至 90% 的端到端可用应用 — Andrei Manolache > [14:29](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=869s) 底层的代码生成可以使用任何通用 LLM，但通用大模型无法内生理解并持久保留欧洲空管深奥的业务约束，必须通过专有的上下文层进行强化。 — Razvan Marginean > [18:24](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=1104s) 通过引入更简便的工具乃至自然语言交互界面，进一步弥合深度技术知识与一线运营知识之间的鸿沟，打造一个真正一体化的产品栈 — Razvan Margauan > [18:24](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=1104s) 这是一场宏大的长期数字化转型工程，绝不可能在一夜之间一蹴而就。 — Razvan Margauan > [19:21](https://www.youtube.com/watch?v=ZnqOx4xmrg8&t=1161s) DesignVerse 的终极目标是赋能客户完全掌控自己的数据资产以及过去 10 到 20 年所积累的业务逻辑。 — Andrei Manolache ## Mentioned EUROCONTROL, DesignVerse, Andrei Manolache, Razvan Marginean, Razvan Margauan, Nancy Scola, FAA, INM (Integrated Network Manager), Eurocontrol Design System, Claude Code, Windsurf, Codex, Cursor, Anthropic, React, Java, Rust ## Video description Andrei Manolache - Chief Executive Officer, DesignVerse Razvan Margauan - Program Manager, EUROCONTROL Nancy Scola - Veteran Technology Reporter, — How do you modernize mission-critical aviation infrastructure without disrupting live operations? EUROCONTROL — the organization coordinating air traffic operations across Europe — shares how it is transforming legacy air traffic systems using DesignVerse, creating a live context layer across architecture, design systems, and operational logic to safely modernize software directly in production. How to turn legacy systems into structured operational context by connecting architecture, design systems, and runtime logic into a unified transformation layer \- Eliminate drift between design and production, with every change aligned to real operational constraints \- Proven modernization in one of the world’s most regulated, safety-critical aviation environments [Source](https://www.youtube.com/watch?v=ZnqOx4xmrg8)

[0:59] Razvan Marginean: 在进行现代化改造的同时，我们绝对不能将系统关闭 [0:59] Razvan Marginean: 每延误一分钟的成本大约是 100 欧元（约合 115 美元） [7:46] Andrei Manolache: 任何人都明白，航空、金融、医疗这类受监管行业中，充斥着服役 10 年、20 年乃至 40、50 年的古老软件。 [11:02] Andrei Manolache: 经过用户验证后，交付给开发团队的是一套完成度高达 80% 至 90% 的端到端可用应用 [14:29] Razvan Marginean: 底层的代码生成可以使用任何通用 LLM，但通用大模型无法内生理解并持久保留欧洲空管深奥的业务约束，必须通过专有的上下文层进行强化。 [18:24] Razvan Margauan: 通过引入更简便的工具乃至自然语言交互界面，进一步弥合深度技术知识与一线运营知识之间的鸿沟，打造一个真正一体化的产品栈 [18:24] Razvan Margauan: 这是一场宏大的长期数字化转型工程，绝不可能在一夜之间一蹴而就。 [19:21] Andrei Manolache: DesignVerse 的终极目标是赋能客户完全掌控自己的数据资产以及过去 10 到 20 年所积累的业务逻辑。

[0:04] Nancy Scola: How do you modernize mission-critical aviation infrastructure without disrupting live operations? [0:04] Razvan Marginean: EUROCONTROL handles approximately 10 million flights annually, with peak traffic reaching up to 37,000 flights per day. [0:50] Razvan Marginean: every single minute of system delay incurs approximately €100 ($115 USD) in cost passed to airlines and ultimately passengers. [6:04] Razvan Marginean: Rather than applying AI directly to mission-critical "crown jewel" flight systems on day one, EUROCONTROL focused on modernizing their broader engineering practices, tooling, and governance. [8:44] Andrei Manolache: The complete front-end and back-end modernization was completed in just 1.5 months. [8:44] Andrei Manolache: The project achieved a 50% reduction in delivery costs and doubled engineering code output (2x productivity) compared to standard coding tools like Claude Code or GPT Codex. [12:03] Andrei Manolache: DesignVerse updates both the front-end UI and the underlying back-end logic simultaneously. [13:16] Razvan Marginean: Specialized tools like DesignVerse provide the persistent aviation domain context, architecture mapping, and UI design system alignment required for regulatory compliance. [17:32] Andrei Manolache: the overall vision for DesignVerse extends well beyond air traffic control; the goal is to become the vertical solution for enterprise software transformation across any heavily regulated sector dealing with legacy infrastructure. [19:21] Andrei Manolache: By targeting model sizes under 100 billion parameters, these models can run locally on consumer-grade or commodity GPU hardware that organizations can easily acquire in bulk, ensuring rapid deployment and absolute data privacy for high-security environments.

[0:04] My name is Nancy Scola and thank you so much for  being here. I'm a journalist based in Washington, [0:08] D.C. And I'm happy to be here this morning to  moderate a conversation on what I think is a [0:12] particularly fascinating topic. How Europe's air  traffic control network, known as Eurocontrol, [0:18] is modernizing using AI. And the specific  question we're going to answer is, [0:22] how do you go about deploying AI that is a perfect  fit for a high-stakes industry like aviation? [0:28] And we have a great set of guests here  joining us to discuss this. We've Andre [0:32] Manolaque. I practiced that. Thank  you. Andre is CEO of Design Verse, [0:37] a startup founded in Bucharest in 2024 that builds  AI software for well-established industries. [0:44] And we have Razvan Marginean. Razvan is an  innovation program manager at Eurocontrol. So, [0:50] Razvan, we're going to start with you and we're  going to start at the beginning. When it comes to [0:53] the AI deployment project we're here to discuss,  what problem were you actually trying to solve? [0:59] Thank you, good morning. Okay, I'm I'm also on.  Um Look, we and I'll give a bit of context about [1:07] Eurocontrol and what we do as well because I  think it's very relevant for the discussion. [1:10] One of the problems that we had to solve is how  to get from a legacy infrastructure built over [1:16] decades of in-house experience into a more modern  infrastructure. And that's part of this whole [1:22] digital transformation that we're undergoing  that's still ongoing now at Eurocontrol. [1:27] Um so, maybe if you want, I could also give a  bit of context on that. Yes, please. Sure. Oh. Um [1:36] I anticipated that maybe not too many people are  familiar with Eurocontrol. Maybe a a quick show of [1:41] hands if you know or have heard of Eurocontrol  before. So, okay. Well, glad to see some [1:48] some hands up. In a nutshell, Eurocontrol  it manages the aviation network. We're not [1:55] air traffic controllers per se. We do network  management. Um basically around 10 million flights [2:01] a year pass through our airspace. Uh and on a  busy day we have around 37,000 flights per day. [2:07] So, I think what makes Eurocontrol unique, uh  it's very similar to FAA uh and it's managing [2:14] a lot of different states, a lot of different  stakeholders so that you have a combined airspace [2:19] and we act as one. Um and that's definitely a  challenge. Um definitely in terms of systems, [2:25] but also in terms of stakeholder management. So,  as I said, um it's quite a legacy organization. [2:34] Um really focused on air traffic control  flight management. And a lot of interconnected [2:40] stakeholders. At the end of the day we make  sure or we try to optimize the minutes of delay. [2:45] Why? Because minutes of delay cost. And how do you  do this? You do this through coordination across [2:50] different airlines, air traffic control centers,  and so on and so forth. So, not not to bore you [2:54] with all the industry jargon. Uh but every flight,  so sometimes when you're flying from the US and [3:00] you have to go through Europe, you may have a slot  that's missed. That's partly because of us. Uh [3:06] fundamentally that's what we do operationally. But  actually we have a lot of uh software competence. [3:11] So, we have a single system that manages  that. Which makes it um a bit of a challenge. [3:18] So, one of our challenges is for example this  regulated environment that we operate in. [3:23] Um keeping know-how and governance over our  stack, over our data, and integrating these [3:31] more modern processes, AI, and so on and so  forth into our existing engineering processes, [3:36] into our tool stack, converting from a legacy  tool stack, which is very niche. It's not a um [3:41] enterprise solution that you get off the shelf.  It's really tailored to our particular needs. [3:46] Uh so, this was a fundamental challenge  that we had of how do we speed up our [3:50] development process? How do we reshape a bit  to access these new technologies on the market? [3:56] And that's where our journey started. Again, for  us, mission critical is a system that if it's [4:04] down, every minute of delay is the equivalent  of okay, 100 euros, but around 115 dollars, [4:10] which is cost passed on to the airlines, which  fundamentally is cost that the end users or the [4:16] flight or the passengers have to undertake. So,  this means that the system is unique. 30 years [4:24] plus of engineering technical debt, and we can't  switch it off, right? So, every time we have to [4:29] modernize, we have to make sure that the existing  system still works. And that's part of this [4:36] digital transformation program, which is  called the INM, Integrated Network Manager. [4:40] And one of the key things that we wanted to do is  to also invest in new technologies in cloud-based [4:45] operations, which is very new for air traffic  control, but also in AI and speeding up our [4:51] existing processes. So, we have two good use cases  that we see in AI. One of them is more speeding up [4:59] our, let's say, requirement gathering process.  So, because we have multiple stakeholders, [5:04] the journey to get from requirement to actual code  is quite lengthy. And we've seen that AI or some [5:12] technologies help us do this much quicker. Which  for us translates into quicker delivery time, [5:17] which fundamentally from a program of such a large  scale, this is one of the number one criteria. [5:22] And the other is transitioning to a more  modern technology stack. We have quite some [5:28] ancient technologies, which are modern, which are  traditional, but very, let's say, safety proof, [5:33] but we have to adapt to a new industry  context. So, one of the challenges that we [5:40] had was to do exactly this. How do we do it  quickly? How do we get to a new tech stack? [5:45] And how do we make sure that uh it all  works seamlessly? So, I think maybe Andre, [5:53] you could uh give us a bit of a use case what we  we've done or Before before we jump to Andre, what [5:59] you mentioned uh the sort of like mission-critical  systems that you were looking to modernize. [6:04] Why I mentioned Design Versus a 2-year-old  company? As great as Andre is, why were [6:08] you willing to trust that modernization to a  company that's only been around a couple years? [6:13] It's a very good question. Um so, we are at the  brink of this digital transformation journey. We [6:20] also had a let's say difficult part of the program  where we needed to deliver much quicker. Um and [6:27] we thought look, now is a good time to invest. If  we need to get uh quicker development capability, [6:33] why not try something new? Of course, we didn't  target our mission-critical our crown jewel, [6:37] so to speak, from the onset, right? This is not  something that's uh easy to do and there's a whole [6:43] certification process. Don't imagine that uh now  you can switch on a tool and suddenly your flight [6:48] management system is changed overnight. It doesn't  work like that. Uh but it gave us an opportunity [6:52] to try to modernize our um whole development  engineering practices so that we can start to [6:58] target these more mission-critical systems once  we have the processes, the governance in place. [7:03] It's not just about a magical tool. It's the  whole ecosystem around that. And we chose uh well, [7:08] one of our partners uh for the fact that it's  something new, it's something that there is not [7:14] uh let's say consistent or I haven't seen a  consistent market offering for these kind of [7:18] technologies. And it was very well uh mapped  onto one of our pain points, which was really uh [7:24] optimizing the time it took us from requirements  to actually getting something to production. [7:29] So, that's what we we started off with and  then we decided to continue to to scale. Okay. [7:33] Andre, you're going to give us a bit of  a demo and overview of what you built. [7:37] Just before we get into that, the when you got  pulled into this project, did you look and say, [7:41] "Okay, this is going to be easy. This is we  know exactly what to do here." Definitely not. [7:46] As everybody in this room can imagine,  regulated industries such as aviation, banking, [7:52] anything along those lines, health care, and  so on. There is software which is 10, 20, 30, [7:56] 40, 50 years old even. So, we knew from the very  beginning that this was going to be a challenge, [8:01] but our mission from the very beginning at Design  Verse was to create a solution, an AI solution [8:07] that can be tailored to critical industries. In  the sense that everybody knows about cloud code, [8:13] Windsurf, Codex, Kiro, and so on, which are  very good for general use cases, or if you want [8:18] to build a 10 people application for your company  where you manage something very simple, but there [8:24] are very few companies who can actually create  something extremely tailored for an industry [8:28] pillar. And that's what we wanted to prove. We  wanted to start with something very, very hard, [8:33] which is aviation, air traffic management, and  then from there to start extending into a lot [8:38] more regulated industries such as banking,  cybersecurity, health care, and and so on. [8:44] Okay. Great. Do you want to give us a little bit  of I think you have an overview for us. Yeah, [8:47] definitely. As Razvan mentioned, we didn't  go into the mission-critical software from [8:51] the very beginning, obviously. So, the very first  modernization project that we were targeted with [8:56] was an application which is still targeted to the  core user base, but it was mainly an information [9:02] board for those users to see all of the flight  details, the slots for the airports, and so on. [9:08] So, the scope was to modernize a Rust plus HTML  CSS application into React plus Java. React [9:15] was chosen because the existing design system,  which is called the Eurocontrol design system, [9:20] was already written into React. It's one of the,  if not the only, aviation-specific design systems. [9:26] So, it was just natural to just reuse  that information where they spent all [9:30] of the resources to build it to modernize  their existing tech stack. And essentially, [9:36] what we saw after the modernization process is  that in 1 and 1/2 months we managed to redo the [9:42] front end and the back end of the application with  a 50% lower delivery cost and two x productivity [9:49] in terms of the engineering code output. This is  being compared with pretty much market standard [9:55] coding tools like cloud code or GPT codex. In this  case specifically it was cloud code and that's [10:01] what we started to see along all of our clients  from all of the business pillars that we were [10:05] targeting. Those benchmarks are done specifically  on very very specific industry standards. [10:11] Nobody cares in enterprise if your SV bench is  the best because you're not doing math problems [10:16] when you're doing aviation obviously. So the  benchmarks are done on those existing code bases. [10:22] What this brings us to is a compounding  effect. The more our users use our platform, [10:27] the more the system learns how they're  building code, how to update it, [10:31] how the code base is going to evolve and so on,  which obviously created compounding effect where [10:36] it understands from an architecture perspective  how the software is being written and how it [10:41] can be translated across all pillars of their  organization. I think you guys have over 5,000 [10:45] employees in total. So it's a pretty massive  scale. It's not the scale of Oracle or IBM, [10:52] but still it's a very very complex organization  that creates those those types of software. [10:57] And essentially from a visual standpoint this  is how it looked before we came in. As you [11:02] can see it's pretty legacy. There isn't like a  cohesive way into how to use the application, [11:07] the UX of it and so on. And after we started the  basically the people started using design verse, [11:14] the end result was something that was actually  usable by the users. Can you I'm so sorry to [11:18] interrupt. Can we can the slide one before exactly  what are we looking at there? Is it So this is an [11:23] application for air traffic controls where they're  seeing the airspaces and the limitations across [11:27] the map of Europe. It's used for information  purposes. Essentially if there is a conflict or [11:33] if there is a sector which has a regulation on top  of it, they use this application to see with that [11:39] the plane basically can pass through that  zone. So, just as a high-level perspective, [11:44] each country is divided into different  airspaces according to the sectors, [11:48] to cantons if you're in Switzerland and so on, and  they can see in here where they can distribute the [11:53] flights across the the countries in Europe. And  if there's something overlapping, if there's a [11:57] conflict spanning across multiple flights and  so on. multiple countries and so on. Okay. [12:03] And essentially, we modernized it, and as you  can see right now it's much more cohesive. So, [12:09] this was done by an internal team, UX designers  and architects, obviously, and the result was a [12:15] fully working application. In our platform, which  is the web app, this is more targeted for product [12:22] use cases like AB testing, stakeholder validation,  and so on. Just with natural language, you can [12:27] change both the front and then the back end at the  same time. The main difference between tools like [12:31] lovable, bold, whatever, Replit, and so on, is  that what you're seeing right now on the screen [12:36] is a fully working end-to-end application hosted  on a server in Brussels, in Belgium. The person [12:41] who put in here a prompt is modifying both the  front and then the back end at the same time. [12:45] The main point being, after you validate this  with the users, you can actually give the full [12:50] end-to-end working application to a developer,  and it's going to be 80 to 90% what they would [12:55] actually put inside production. Sure, you run QA,  you put it into CI/CD, you put it into staging [13:00] and then production, but it's much more closer  than somebody working in cloud code, and then [13:04] trying to stitch it together somewhat what the  product designers are doing with with the users. [13:13] Okay. That's pretty much it. Yeah. Thank you  very much. So, that what we were just looking at, [13:16] that product, that reflects a detailed  understanding of air traffic control systems, [13:22] which I'm guessing maybe you didn't come into  the project with. How do you make sure that [13:26] your controls expertise is reflected in actually  what ends up being built? So, definitely there's [13:34] a lot of operational expertise that's needed to  understand and to validate this. And this is one [13:39] of the challenges actually that we have a lot  of operational stakeholders without a technical [13:42] background. And we have a lot of technical  know-how without operational understanding. [13:47] So, it's not a it's not so much an easy e-commerce  app that you can kind of pick up on the fly. It's [13:52] really specific industry know-how. Uh which is why  I personally like this type of use case because [13:58] it brings those two stakeholders together. So,  it doesn't separate. Uh and we need to keep our [14:03] operational staff, also our technical staff  because it's years of accumulated know-how. [14:08] Uh still capable to manage such systems. So,  by offering this kind of also visual type of [14:15] approach, this product layer, it makes those two  stakeholders work much more seamlessly together. [14:22] Uh which I think is one of the reasons that we  got this uh much quicker delivery time. Um so [14:29] Yeah, Andre, the same question for you. How did  that work in practice? What was the partnership [14:32] like that you made sure you were tapping into the  expertise that Eurocontrol had? From our product [14:38] design background, we worked mainly in product  and enterprise software before we even started [14:41] Design Verse. We knew from the very beginning  that in order for something to be successful, [14:45] so not just a POC that fizzles out after a month,  you need to understand from the deepest level how [14:50] the software works for your client. And we've  been very very close with their team from the [14:54] very beginning inherently even before our system  started to understand how they write software, [14:59] we needed to understand what to tell  the system, what to tell our agents, [15:03] and all the whole AI pipeline to basically try to  comprehend all of their years of engineering. So, [15:11] we've been very close from the beginning. After  we started understanding how the software works, [15:14] the architecture, and how to integrate to the  deepest level with their existing tech stack, [15:19] then we started uh procuring this information  inside the system, and then it started becoming [15:24] more and more automated along the way. But, it has  been a very very tight collaboration from the very [15:28] beginning. Okay. Uh Razvan, one of the challenges  seems to be rather than Andre talks about the [15:33] benefits of not using Cloud Code or some of these  other off-the-shelf models is that you have a [15:38] something that is we talked about reflects the  expertise of the industry that you're working in. [15:41] It also one of the benefits of a working with a  model like Cloud Code is you know it's going to be [15:45] updated. You know there's a big company behind it  who's putting a lot of money into its development. [15:49] Are there any tradeoffs with going this  really custom-built uh model approach? [15:54] I think um so maybe to to take a step back, our  AI landscape is a bit diverse. It's not that we [16:01] we found one company and we said, "Okay,  everything is going to be done with this." [16:04] Uh a lot of our investments of okay, in a  nutshell, we set up what we call an AI center [16:09] of excellence where we look at the different AI  initiatives across the organization. So, this one [16:15] uh as Andre also mentioned, it's really it's nice  because it bridges that gap between product design [16:21] uh that operational and technical know-how. Um  as far as I know, I think the model is agnostic [16:26] of so I think you can use any LLM uh for the code  generation part. But, what is critical for us is [16:33] really that aviation context which uh Cloud can't  pick up so easily or retain to that level. I think [16:39] Cloud is more a general-purpose solution which  we rely quite heavily on to augment our software [16:44] development capabilities, but I don't see it uh  necessarily generating a product based on these [16:50] requirements within our specific technical  landscape. So, we we combine the two. Um [16:57] but again, we we don't have a single uh let's  say investment approach where we said, "Okay, [17:02] now we have this magical tool and it solves all  our problems." We look at the whole landscape [17:06] holistically, and based on our business  challenges, so there are many. Uh again, [17:11] it's legacy, it's 30-plus years uh of know-how  uh which now needs to modernize. Also, going [17:17] cloud-based introduces a different uh type of  skill set that we didn't have internally before. [17:22] Uh, so there are a lot of things that we have to  manage in parallel. Okay. Okay, thank you. Andre, [17:26] what happens next? You take this what you've  built here and sell it to the FAA, um, which [17:32] needs the help, uh, other aviation control systems  around the world, or can you take this and apply [17:37] this to industries beyond aviation? Uh, it's a  combination of both. Obviously, with the knowledge [17:43] that we have for a very specific business pillar,  we're trying to extend that to other clients who [17:46] need this help as well. And that's why even with  Eurocontrol, besides being a client, we're also [17:51] working with them to spread this information  about what they built for the past 20 years, [17:55] and what can be applied to other clients in the  same sector, irrespective if there's regulators or [17:59] just regular ANSPs who are controlling the local  traffic for their country. But, uh, the goal of [18:05] the company is to become the vertical solution for  any enterprise use case that we target. Currently, [18:12] it's aviation, banking, and cybersecurity.  These are the main pillars that we operate in, [18:15] but in the future, this can be tailored  to pretty much any industry that involves [18:20] software development and legacy applications.  Okay. Razvan, you mentioned that you started [18:24] with a bit of a small project, right? Sort of  proof of concept, and you're going on from there. [18:27] What is the big project that you're thinking of  in 10 years? Oh my gosh, if I can pull this off, [18:31] it's going to be very exciting. I think, uh,  without giving away too much of of our strategy, [18:37] um, look, I think it would be nice. So, we  have a lot of, um, stakeholder-facing products [18:44] uh, for our operational clients. It would be  nice if we could harmonize and we could build, [18:49] uh, on this product stack, uh, using a simpler  tool, using, uh, let's say, even some more [18:55] natural language, so that we further bridge,  uh, this gap between a lot of technical know-how [19:00] and a lot of operational know-how, and we have  an integrated product stack. So, that's something [19:04] that I would like to see scale, uh, going forward.  Again, it's a large transformation project, [19:10] so that imagine it's going to happen overnight,  but now we're managing to put the frameworks [19:15] so that when we can start to scale, at least it  meets our requirements. Okay, excellent. Andrei, [19:21] I I'm mostly a political reporter, so I'm going  to take the prerogative to ask a political some [19:25] something of political question. You've talked a  little bit or written a little bit about the idea [19:29] that building these custom-built models, there's a  little bit of a digital sovereignty aspect of this [19:33] that you're not so dependent on a Anthropic model  that the US government might decide to put export [19:38] controls on. Can you just talk a little give a  little bit of a pitch on that? Yeah, definitely. [19:43] Our aim is to empower our clients to take control  over the data and everything that they built for [19:48] the past 10 to 20 years. And essentially, what  we're building towards is our own LLMs which can [19:53] be deployed on premise, which are much smaller  than Claude GPT, GLM, Kimiko Tree, and so on, [19:59] because what we've seen is that with the right  context and the right guidelines, essentially, [20:04] you can use much smaller models to write code  much more efficiently than just using Claude to do [20:09] very, very, very specific business tasks. And what  we're building towards is a solution which can be [20:14] deployed very easily on premise. We're trying to  get our models to be under 100 billion parameters, [20:18] meaning that you can put it on a consumer grade  GPU, and this can be acquired very easily in bulk [20:23] from any supplier that you want, and it can  become a solution that can be deployed very, [20:28] very rapidly across clients with very, very  sensitive information. Okay, excellent. Thank you, [20:32] Andrei. Thank you, Razvan. Thank you all for  being here. Appreciate it. Thank you. [applause]