CIRCE Deploys Self-Tuning AI Agents to Multiply Proposal Pipeline

Bottom-up Innovation: Boosting Collaborative Product Development with AI — Ai4

Ai4

Deploying evolutionary metaprompting across internal multi-agent workflows eliminates large language model variability, allowing research organizations to scale contract bidding and revenue without heavy capital spending.

AI-assisted summary of Bottom-up Innovation: Boosting Collaborative Product Development with AI — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by Ai4 on August 26, 2025, this presentation features Miguel Ángel Castán Lascorz and Luis Luengo from CIRCE Technology Center. The session details how CIRCE applied artificial intelligence internally through an open innovation challenge called the AI Olympics, highlighting their proposal-evaluating assistant, La Ofertadora.

## TL;DR To free technical staff from tedious administrative proposal tasks, CIRCE Technology Center organized a four-month internal competition called the AI Olympics, where cross-functional teams developed practical AI prototypes. The top solution, La Ofertadora, automates RFP web scraping, document analysis, internal capacity matching, and pricing. To resolve nondeterministic variability across multi-agent pipelines without sacrificing contextual inference, the team developed an evolutionary metaprompting algorithm that iteratively optimizes prompts for stability and quality. The deployment scaled CIRCE's detected tender opportunities from dozens to over 3,000, establishing a replicable blueprint for low-cost, bottom-up AI adoption.

## Atmosphere & style The presentation carries an energetic, pragmatic, and collaborative vibe, seamlessly blending local pride in Zaragoza's tech ecosystem with real-world AI engineering challenges. Castán Lascorz and Luengo deliver a polished, accessible pitch that balances technical depth—such as agent orchestrations and evolutionary Pareto frontiers—with relatable workplace frustrations. The session feels both grounded and inspiring, demonstrating how open internal experimentation can ignite genuine organizational transformation without massive capital expenditure.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=4s) Manual RFP tracking previously consumed excessive engineering time, prompting CIRCE to automate opportunity discovery and let technical staff focus on core sustainability projects. - [01:36](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=96s) CIRCE Technology Center drives European sustainability initiatives by uniting multidisciplinary teams across energy, digitalization, and environmental sciences. - [04:22](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=262s) The internal 'AI Olympics' established a four-month bottom-up open innovation framework, giving employees dedicated company time to develop AI solutions for real operational needs. - [04:22](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=262s) The competition yielded diverse AI prototypes, addressing challenges such as vineyard irrigation optimization, bird migration mapping for wind farms, and smart traffic signal control. - [07:22](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=442s) Top-performing tool 'La Ofertadora' was built as a multi-agent pipeline automating tender scraping, document analysis, internal skill matching, and pricing calculations. - [07:22](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=442s) Production deployment exposed a fundamental LLM consistency dilemma: raw models introduced cascading non-deterministic errors, while rigid hyperparameter caps destroyed reasoning capability. - [10:52](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=652s) CIRCE resolved agent inconsistency using Evolutive Metaprompting, an approach where meta-agents iteratively mutate and score prompt variations against quality and stability benchmarks. - [10:52](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=652s) Evolutionary prompt optimization preserves white-box transparency, generating inspectable natural language instructions that automatically adapt as underlying AI models evolve. - [15:07](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=907s) The operationalization of La Ofertadora expanded CIRCE's identified opportunities from under 100 to over 3,000, projecting a 9x revenue increase. - [15:07](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=907s) Transformative organizational AI adoption does not demand massive capital expenditure, but rather strategic agility, bottom-up empowerment, and a culture of experimentation.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=4s) Rethinking How Time Is Spent on Green Projects

Miguel Ángel Castán Lascorz and Luis Luengo open the presentation by asking the audience a central question: "How much time do you spend thinking green?" As project managers at CIRCE Technology Center in Spain specializing in energy flexibility, they realized that an enormous portion of their working hours was consumed by hunting down potential opportunities, reviewing fragmented sources of information, and determining whether incoming Requests for Proposals (RFPs) fit the capabilities of their multidisciplinary teams. This left far too little time to actually execute the green, sustainable projects they cared about.

Last year, CIRCE identified 22 interesting RFPs manually, spending an average of eight minutes analyzing each document. To eliminate this bottleneck, they created **La Ofertadora** (the "RFPs analyzer"). Since its rollout, the system identified 131 relevant opportunities in just one minute via an automated, single-click workflow, freeing up technical staff to focus on substantive sustainability work.

- **Manual baseline:** 22 RFPs identified last year at an average of 8 minutes spent analyzing each. - **La Ofertadora efficiency:** 131 RFPs identified in 1 minute with a single click.

---

[01:36](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=96s) Zaragoza and CIRCE Technology Center

The presenters welcome the audience to Zaragoza, Spain. Acknowledging common stereotypes of Spain—paella, sangria, and beaches—they playfully clarify that Zaragoza is known for different landmarks, notably the Basilica del Pilar and the Ebro River. Connected deeply to water since ancient Roman times, Zaragoza hosted the World Expo in 2008. Today, the city is in the midst of a modern technological revolution and is slated to host the largest data center in Southern Europe.

CIRCE Technology Center embodies this regional transformation. Its mission is to convert environmental challenges into innovation opportunities across three pillars: energy, sustainability, and digitalization. CIRCE employs over 300 professionals spanning diverse disciplines—including chemists, architects, data scientists, software developers, and laboratory technicians—driven by the belief that high-impact solutions arise when cross-disciplinary teams collaborate.

Beyond internal software, CIRCE leads several prominent European technological initiatives: - **REDOL:** Leverages AI, robotics, and advanced recycling to recover materials from urban waste, supporting cities in achieving climate neutrality by 2040. - **ELECTRA:** Electrifies high-temperature industrial furnaces using electrode, induction heating, and microwave technologies. - **EFFORT:** Strengthens the cyber resilience of energy infrastructure using AI and self-healing systems to mitigate risks and speed up disaster recovery. - **INCIT-EV:** Develops Europe's first dynamic wireless charging lane for electric vehicles.

---

[04:22](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=262s) Bottom-Up Innovation: The AI Olympics

Reflecting on their client-facing R&D work, CIRCE's leadership and staff asked: "What if we apply the same innovation that we provide to our clients, to ourselves?" and "What if we can exploit the full power of AI into our multidisciplinary projects?" To answer how to operationalize this, CIRCE's board launched the **AI Olympics**—a four-month internal bottom-up open innovation competition.

Teams were given dedicated company time to collaborate across departments and build AI-powered prototypes targeting real operational and market challenges. This framework produced multiple validated solutions: - **BIVAR:** A vineyard irrigation optimization mobile app that predicts water needs via deep learning and delivers irrigation recommendations to growers over WhatsApp, delivering up to 20% water savings without compromising grape quality. - **ONGUARD:** Predicts shifts in bird populations in response to climate change up to 25 years into the future, providing distribution maps so wind farm planners can avoid high-impact habitats and prevent regulatory conflicts. - **VALTIR:** Evaluates the State of Health (SoH) of electric vehicle batteries simply by plugging the vehicle into a smart charger. - **DRAGON:** Dynamically optimizes the real-time capacity and throughput of power transmission lines. - **SAFETRA:** Uses computer vision at pedestrian crossings to infer crossing intent and dynamically modulate traffic lights, cutting traffic congestion and saving an estimated 10 tons of $\text{CO}_2$ per traffic light point each month.

---

[07:22](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=442s) La Ofertadora Architecture & The LLM Consistency Problem

La Ofertadora emerged as the overall champion of the AI Olympics. Designed to streamline technical proposal development, the tool crawls public tender sites, analyzes extensive specifications, matches task requirements against internal staff skills, and calculates competitive pricing.

``` +--------------------------------------------------------------------+ | Multi-Agent System | | | | +------------------+ +------------------+ | | | Web Scraping & | --> | Doc Analysis & | | | | Auto-Download | | Info Extraction | | | +------------------+ +--------+---------+ | | | | | v | | +------------------+ +------------------+ | | | Pricing Context | <-- | Capacity Matcher | | | | & Estimation | | (Team Assignment)| | | +--------+---------+ +------------------+ | | | | | +------------------------+ | | v | | +-----------------------+ | | | Orchestrator Pipeline | | | +-----------+-----------+ | +------------------------------------|-------------------------------+ v +-------------------------+ | Final Evaluation Report | +-------------------------+ ```

Despite initial success, the team ran into a major production hurdle common to large language models (LLMs): nondeterministic variability. When evaluating identical proposal requirements, the assignment agent would intermittently change team alignment scores or suggest entirely different technical units. This variability cascaded across downstream agents, leading to extreme output uncertainty.

While lowering temperature or tuning decoding hyperparameters forced determinism, it made the models rigid and destroyed the "spark of inference" necessary for contextual synthesis. The presenters emphasized the classic dilemma: "We cannot kill creativity, but we need to guide it." Manually engineering prompts for every agent combination was impractical, driving them to find an automated approach.

---

[10:52](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=652s) Evolutionary Metaprompting

Inspired by a *Nature* publication, CIRCE designed **Evolutive Metaprompting**, using AI agents to iteratively evaluate and rewrite the prompts of other operational agents.

``` +--------------------------------------------------------------------+ | Evolutive Metaprompting Loop | | | | +--------------------------------------------------------+ | | | Initial Population: Prompt Variations & Hyperparams | | | +---------------------------+----------------------------+ | | | | | v | | +--------------------------------------------------------+ | | | Multi-Objective Evaluation: Consistency vs Quality | | | +---------------------------+----------------------------+ | | | | | v | | +--------------------------------------------------------+ | | | Selection: Identify Pareto-Optimal Prompt Candidates | | | +---------------------------+----------------------------+ | | | | | v | | +--------------------------------------------------------+ | | | Meta-Prompt Agent: Mutation & Crossover of Prompts | | | +---------------------------+----------------------------+ | | | | | +--- (Repeat for N Generations) | +--------------------------------------------------------------------+ ```

The system pairs operational agents (e.g., the team-assignment agent) with a meta-agent acting as an automated prompt engineer. Starting from an initial population of prompt structures and parameters, candidates are tested against historical proposal benchmarks and scored on both output quality and consistency.

Top-performing instructions undergo evolutionary crossover and mutation, guided by the meta-agent. Over successive generations, prompts migrate toward the Pareto-optimal top-right quadrant (high accuracy, high stability). The resulting instructions abandoned brittle, rigid rule lists in favor of balanced prompt structures that preserve reasoning flexibility while enforcing strict deterministic constraints.

Key benefits of this approach include: - **Adaptive automation:** Continuously optimizes prompts even as underlying models, APIs, or data schemas evolve. - **System-wide robustness:** Prevents cascading errors across multi-agent pipelines. - **White-box transparency:** Output prompts remain natural text that engineers can inspect, audit, and supervise.

---

[15:07](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=907s) Business Impact and Cultural Transformation

The deployment of La Ofertadora drastically expanded CIRCE's commercial and grant pipeline: - **Opportunities detected:** Grew from 86 manually identified opportunities in the prior year to more than 3,000 using La Ofertadora. - **Proposal submissions:** Expanded from 45 submitted proposals to an addressable contribution capacity of approximately 1,600. - **Project wins:** Increased the potential win volume from 5 historical victories up to nearly 200 successful bids. - **Financial and organizational growth:** Projections indicate a 9x increase in annual line-of-business revenue and the creation of roughly 60 new technical jobs.

Luis and Miguel Ángel conclude by underscoring that applying AI internally proved that transformative innovation does not require massive capital expenditure—it requires "vision, agility, and a culture open to experimentation and learning." The AI Olympics evolved beyond a internal competition into an organizational catalyst, demonstrating that "when a technology center turns its intelligence inward, it unlocks a multiplier effect" across technological, human, and environmental domains.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=4s) How much time do you spend thinking green? — Miguel Ángel Castán Lascorz

> What happens when a technology center decides to apply artificial intelligence... to itself?

> [04:22](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=262s) What if we apply the same innovation that we provide to our clients, to ourselves?

> [04:22](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=262s) What if we can exploit the full power of AI into our multidisciplinary projects?

> [07:22](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=442s) We cannot kill creativity, but we need to guide it.

> [15:07](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=907s) vision, agility, and a culture open to experimentation and learning.

> [15:07](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=907s) when a technology center turns its intelligence inward, it unlocks a multiplier effect

> a replicable model of bottom-up innovation that turns AI into a catalyst for real change.

## Mentioned Miguel Ángel Castán Lascorz, Luis Luengo, CIRCE Technology Center, Ai4, La Ofertadora, AI Olympics, Zaragoza, Basilica del Pilar, Ebro River, World Expo 2008, REDOL, ELECTRA, EFFORT, INCIT-EV, BIVAR, ONGUARD, VALTIR, DRAGON, SAFETRA, Nature, Evolutive Metaprompting

## Video description Miguel Angel Castan Lascorz, Project Manager in AI-based solutions for energy efficiency and sustainability, CIRCE Technology Center Luis Luengo, Project Manager, CIRCE Technology Center

Abstracts

What happens when a technology center decides to apply artificial intelligence... to itself? In this talk, I’ll share how CIRCE launched its first internal AI Olympics—an open innovation challenge where multidisciplinary teams developed AI-based solutions for real organizational needs. I’ll walk you through the process, the outcomes, and our most successful solution: La Ofertadora, an AI-powered assistant that’s transforming how we prepare R&D proposals. You’ll learn how AI became a practical, strategic tool within our daily workflows—without requiring massive investment. Join us to discover a replicable model of bottom-up innovation that turns AI into a catalyst for real change.

[Source](https://www.youtube.com/watch?v=hgM4WSIfdE8)

中文简报

CIRCE部署自调优AI智能体成倍扩充提案管线 在内部多智能体工作流中应用进化元提示词消除了大语言模型的不确定性，使科研机构无需巨额资本支出即可扩大合同竞标规模并增加收入。 为将技术人员从繁琐的提案行政事务中解放出来，CIRCE技术中心举办了为期四个月的内部竞赛“AI奥林匹克”，由跨职能团队开发实用的AI原型。其中表现最优的方案La Ofertadora实现了招标书（RFP）网络抓取、文档分析、内部能力匹配和定价的自动化。针对多智能体管线中的非确定性波动，团队在不牺牲上下文推理能力的前提下，开发了一种进化元提示词算法来迭代优化提示词的稳定性和质量。该系统的部署使CIRCE发现的招标机会从数十个跃升至3,000多个，为低成本、自下而上落地AI提供了可复制的范式。 西班牙CIRCE技术中心发起为期四个月的内部“AI奥林匹克”开放创新挑战赛，为跨学科员工提供专属工作时间，针对实际运营与环保痛点研发出葡萄园智能灌溉、风电场鸟类迁徙预测等多款原型应用。 (4:22) 夺冠项目“La Ofertadora”构建了涵盖网页抓取、文档分析、内部技能匹配与定价测算的多Agent流水线，大幅消除手动追踪RFP的流程瓶颈，实现一键1分钟内筛出131个匹配项目。 (7:22) 针对生产环境中大语言模型输出不确定导致的级联误差，团队采用演化元提示（Evolutive Metaprompting）机制，借助元Agent迭代变异与筛选提示词，在维持白盒透明度和推理能力的前提下兼顾输出稳定性。 (10:52) La Ofertadora的上线将CIRCE识别的业务机会从过去的86项提升至3000项以上，预计将使相关业务线年收入增长9倍，并创造约60个新的技术就业岗位。 (15:07)

Reading summary

CIRCE Technology Center project managers Miguel Ángel Castán Lascorz and Luis Luengo argue that internal, bottom-up innovation competitions can automate complex organizational workflows and expand business capacity without massive capital expenditure. Through their internal competition, the center developed La Ofertadora, a multi-agent system that streamlines proposal discovery while resolving large language model consistency issues through evolutionary prompt optimization. Fostering Internal Innovation Through the AI Olympics (4:22) CIRCE Technology Center, an applied research organization based in Zaragoza, Spain, launched the "AI Olympics"—a four-month internal competition providing its 300 multidisciplinary professionals with dedicated time to develop AI solutions for operational and environmental challenges. Rather than limiting innovation to external client projects, the initiative encouraged cross-departmental teams of chemists, architects, data scientists, and engineers to build targeted prototypes. The competition produced multiple sustainability-oriented applications: BIVAR, an irrigation optimization app using deep learning to recommend watering schedules over WhatsApp for up to 20% water savings; ONGUARD, a predictive tool forecasting bird population shifts 25 years into the future to help wind farm developers avoid ecological conflicts; and SAFETRA, a computer-vision traffic system estimated to cut 10 tons of CO2 per traffic light monthly by inferring pedestrian intent. Multi-Agent Automation of Proposal Workflows (0:04) The overall winner of the competition was La Ofertadora, an automated tool engineered to remove administrative bottlenecks in reviewing Requests for Proposals (RFPs). Previously, project managers manually located 22 relevant RFPs a year, spending an average of eight minutes analyzing each document. In contrast, La Ofertadora identified 131 qualified opportunities within one minute through a single-click workflow. The system operates via an orchestrated multi-agent pipeline. Automated agents scrape public tender portals, extract document specifications, match required tasks against staff skills to assemble technical teams, and calculate competitive pricing before compiling a final evaluation report. Evolutionary Metaprompting for LLM Consistency (10:52) During deployment, the engineering team faced severe nondeterministic variability from large language models. The capacity-matching agent frequently suggested divergent teams or differing alignment scores for identical RFP texts, creating cascading failures across downstream pricing and synthesis agents. Lowering decoding temperature eliminated necessary contextual reasoning, while manually engineering prompt combinations across multiple agents was unfeasible. To solve this, CIRCE implemented "Evolutive Metaprompting," an automated framework inspired by research published in Nature. A meta-agent iteratively tests an evolving population of prompt variations against historical proposal benchmarks, evaluating candidates for both accuracy and stability. Over successive generations of mutation and crossover, prompts move toward Pareto-optimal performance, establishing consistent outputs without resorting to brittle, overly restrictive rules or obscuring human-auditable prompt text. Pipeline Expansion and Business Growth Projections (15:07) Deploying the multi-agent system significantly widened CIRCE's commercial and grant pipeline. The center reported expanding its annual identified opportunities from 86 manual leads to over 3,000, while expanding its addressable submission capacity from 45 historic proposals to approximately 1,600. Based on these pipeline improvements, CIRCE projects its historical win volume of 5 projects could rise to nearly 200 successful awards. The presenters anticipate this automated intake will drive a 9x increase in annual line-of-business revenue and support roughly 60 new technical hires. The presenters conclude that transformative AI adoption depends more on organizational agility, multidisciplinary collaboration, and experimentation than on significant capital investment. However, the reported ninefold revenue surge and 200 project victories remain forward-looking projections that depend on sustained win rates and operational capacity to deliver on award volumes.

精炼阅读摘要

西班牙CIRCE技术中心通过内部自下而上的“AI奥林匹克”创新竞赛，将多学科人工智能能力转向内部流程再造，开发出自动化提案分析系统La Ofertadora，并利用进化元提示词技术解决大语言模型输出不稳定的难题，显著提高了绿色项目的获取效率与业务增长潜力。 自下而上的AI创新机制与绿色项目研发 (4:22) CIRCE技术中心拥有300余名跨学科专业人员，涵盖化学家、建筑师、数据科学家及软件开发人员，主导了涵盖城市固废资源回收（REDOL）、工业高温窑炉电气化（ELECTRA）、能源基础设施网络韧性（EFFORT）以及电动汽车动态无线充电车道（INCIT-EV）等多项欧洲技术倡议。然而，项目管理人员Luis Luengo与Miguel Ángel Castán Lascorz指出，团队过去耗费大量时间从零散渠道寻找商机并人工评估招标书（RFP），挤占了推进核心绿色项目的精力。为此，管理层发起了为期四个月的全员“AI奥林匹克”竞赛，鼓励跨部门团队利用工作时间构建面向实际业务痛点的AI原型。 该竞赛孵化了多个面向可持续发展的落地技术：BIVAR利用深度学习预测葡萄园用水需求，通过WhatsApp向种植者推送灌溉建议，在保障品质的前提下节水达20%；ONGUARD可模拟长达25年的气候变化对鸟类分布的影响，协助风电规划避开敏感生态区；VALTIR支持在接入智能充电桩时即时评估电动车电池健康状态（SoH）；DRAGON实现了输电线路容量与吞吐量的动态实时优化；SAFETRA则通过计算机视觉识别人行横道行人的过街意向并动态调控交通灯，据估计每个信号点每月可减少约10吨二氧化碳排放。 La Ofertadora多智能体架构与一致性难题 (7:22) 作为“AI奥林匹克”的总冠军项目，La Ofertadora旨在解决技术提案编写的流程瓶颈。此前人工模式下，CIRCE去年仅甄别出22份目标RFP，平均每份耗时8分钟进行分析；而La Ofertadora上线后，通过一键式自动化流水线仅需1分钟即可识别出131个相关机会。该系统采用多智能体协同架构，涵盖网络爬虫与自动下载、文档分析与信息提取、团队能力与人员匹配、定价上下文与成本估算，最终由编排流水线整合输出评估报告。 但在实际落地过程中，团队遭遇了大语言模型普遍存在的非确定性问题。针对完全相同的招标需求，人员匹配智能体会间歇性给出不同的团队匹配得分或推荐不同的技术部门，导致下游智能体产生级联误差，输出结果可靠性大打折扣。演讲者指出，单纯调低模型温度或调整解码超参数虽然能强制输出确定性结果，却会使模型陷入机械僵化，丧失上下文综合与推理能力。如何在保留模型推理灵活性的同时实施严格约束，构成了多智能体协作的核心技术阻碍。 基于进化元提示词的自适应工程突破 (10:52) 受《自然》（Nature）期刊相关论文启发，CIRCE设计了“进化元提示词”（Evolutive Metaprompting）机制，利用上层元智能体对业务执行智能体的提示词进行迭代评估与自动化重写。该流程首先建立包含多种提示词结构与超参数的初始种群，依据历史提案基准库，围绕输出质量与一致性两个维度开展多目标测试评估。 在每一代评估中，处于帕累托最优候选区（兼顾高准确度与高稳定性）的表现优异者会被保留，并由元智能体进行交叉重组与变异演变。随着多代迭代，系统逐步收敛至最优提示词结构，替代了脆弱死板的人工规则列表。该方案具备三大关键优势：其一是自适应自动化，底层模型、API或数据格式变动时系统可自主重构提示词；其二是全局鲁棒性，有效切断了多智能体管道中的级联误差；其三是白盒透明度，优化生成的最终提示词依然保持自然语言形态，便于技术人员进行审查、审计与人工监督。 商机拓展成效与组织长期增长预测 (15:07) 随着La Ofertadora的全面部署，CIRCE的商业与研发项目获取能力实现了量级突破。在实际指标方面，机构识别的项目机会从上一年度人工筛查的86项激增至3000项以上；投标方案的承接与提交能力从原先的45份扩展至约1600份；项目中标预期也从历史上的5次提升至近200个竞标成功预期。 针对长期组织与财务回报，演讲者给出了预测判断：该业务板块的年营收预计将实现9倍增长，并有望在此过程中创造约60个新的技术岗位。Luis Luengo与Miguel Ángel Castán Lascorz总结认为，这一成果表明企业内部AI创新并不依赖巨额资本投入，关键在于敏捷度、战略眼光以及鼓励试验的开放文化；当技术中心把智能技术反哺于自身流程时，将产生跨越技术、人力与环境层面的乘数效应。 演讲者所陈述的业务指标展现了多智能体流水线在精简科研申报管理上的切实收益，但有关年收入增长9倍以及新增60个技术工作岗位的表述属于前瞻性商业预测，其最终落地成效仍取决于后续竞标转化率与项目实际运营数据的检验。

Detailed notes — zh

## Background 本内容由 YouTube 频道 Ai4 于 2025 年 8 月 26 日发布，由西班牙 CIRCE 技术中心的项目经理 Miguel Angel Castan Lascorz 与 Luis Luengo 共同主讲。视频围绕该技术中心如何通过内部自下而上的“AI 奥运会”创新挑战赛，将人工智能技术反哺自身业务，并重点介绍了其获奖项目“La Ofertadora”多智能体系统的落地实践与成效。 ## TL;DR CIRCE 技术中心为解决科研与商业竞标中繁琐耗时的痛点，在内部发起了为期 4 个月的“AI 奥运会”，孕育出专注于自动化 RFP 筛选与方案推荐的冠军工具“La Ofertadora”。面对多智能体系统在实际落地中的随机性与不稳定性，团队引入了基于演化计算的“演化元提示词（Evolutive Metaprompting）”机制，通过 AI 优化 AI 实现了底层推理逻辑的质变与高鲁棒性。该系统的成功应用不仅使商机挖掘量从 86 个爆发式增长至 3,000+ 个、预计带来 9 倍收入增长并创造 60 个新岗位，更展示了一种无需巨额资金、以组织内部赋能为核心的自下而上 AI 创新范式。 ## Atmosphere & style 整场演讲充满热情、务实且极富感染力，两位讲者配合默契，将技术干货与生动的组织变革故事紧密交织。叙事从日常琐碎案头工作的痛点切入，逐步推进到攻克大模型随机性难题的硬核演化算法，最后落脚于令人振奋的业务爆发与组织文化觉醒，情绪曲线层层递进。聆听过程中既能感受到工程师攻克技术顽疾时的严谨与兴奋，又能体会到一家跨学科研究机构勇于“向内动刀”、拥抱自我革新的蓬勃生命力。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=4s) 科研与工程团队面临大量商机筛选与 RFP 评估等案头工作痛点，耗费了本应投入核心绿色研发的宝贵精力。 - [01:36](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=96s) CIRCE 技术中心汇集 300 多位跨学科专家，在能源、可持续性与数字化领域推进多项欧洲重点前沿项目。 - [04:22](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=262s) CIRCE 举办首届内部“AI 奥运会”，打破部门壁垒自下而上孵化出包括葡萄园节水系统 B-VI 在内的多项落地应用。 - [07:22](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=442s) 冠军项目 La Ofertadora 依托多智能体架构与调度中枢，实现从招标抓取、团队能力匹配到方案报价的全闭环自动化。 - [08:34](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=514s) 为解决大模型在多 Agent 协同中的随机性与输出不一致问题，团队创新性提出了“演化元提示词（Evolutive Metaprompting）”机制。 - [12:21](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=741s) 演化计算框架在一致性与输出质量两个维度进行多目标评估，通过模拟遗传算法中的交叉与变异自动化迭代 Agent 提示词。 - [13:28](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=808s) 演化提示词优化从根本上提升了系统的底层推理逻辑，具备高自适应性、高鲁棒性以及基于文本的完全可解释性三大优势。 - [15:07](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=907s) La Ofertadora 助力商机发现量从 86 个激增至 3000 多个，预计带来 9 倍收入增长并直接带动 60 个新就业岗位。 - [16:03](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=963s) 组织内部的 AI 创新无需巨额资金投入，关键在于将 AI“向内应用”赋能自身，实现技术与团队潜能的乘数效应。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=4s) 绿色转型的效率痛点与“La Ofertadora”的诞生 来自西班牙 CIRCE 技术中心的项目经理 Miguel Angel 和 Luis 登台分享了他们在能源灵活性领域的创新探索。演讲开篇提出了一个引发深思的问题：“你每天花多少时间进行绿色思考？”在日常工作中，团队意识到大量宝贵的时间并非用于执行可持续性项目，而是被无休止的前期事务所消耗——在海量信息渠道中搜寻潜在商机、筛选符合公司技术方向的招标项目（RFP），并反复评估这些项目与内部团队的匹配度。 为了解决这一痛点，团队研发了一款名为“La Ofertadora”（意为“智能标书与商机分析助手”）的 AI 工具，彻底重塑了机会挖掘流程。过去一年中，人工团队仅筛选出 22 项相关 RFP，且每份平均需耗费 8 分钟进行分析；而在 La Ofertadora 上线后，系统在短短 1 分钟内就自动化筛选并分析了 131 项 RFP。这种一键式全流程自动化让团队得以从繁琐的案头调研中解放出来，将核心精力重新投入到真正推动绿色低碳落地的实质性研发项目中。 - **效率对比突破**：过去人工筛选 22 个 RFP，每项平均耗时 8 分钟；AI 助手上线后仅需 1 分钟即可处理并挖掘 131 个商机。 - **核心定位**：将商机识别与技术适配过程从繁重的人工筛查转变为“一键自动化”的智能化工作流。 --- ### [01:36](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=96s) 萨拉戈萨与 CIRCE：从传统名城到跨学科技术枢纽 介绍完项目初衷后，讲者将视线引向他们的所在地——西班牙萨拉戈萨（Zaragoza）。萨拉戈萨并非只有外界普遍印象中的海鲜饭、桑格利亚酒或海滩，这座城市坐拥宏伟的皮拉尔圣母圣殿（El Pilar）与埃布罗河（Ebro River），自古罗马时期便与水资源紧密相连，更曾于 2008 年举办以水为主题的世博会。如今，萨拉戈萨正迎来一场深刻的数字化与科技革命，南部欧洲最大的数据中心也即将在此落户。 作为这一科技转型浪潮的缩影，CIRCE 技术中心专注于能源（Energy）、可持续性（Sustainability）和数字化（Digitalization）三大支柱领域，致力于帮助各行业企业提升竞争力。CIRCE 汇集了 300 多位来自不同学科的专家，涵盖化学家、建筑师、数据科学家、软件工程师及实验室技术员等多元背景。讲者列举了中心目前主导的四项国际重点项目，展现其跨学科技术实力： - **REDOL**：融合 AI、机器人技术与先进循环回收工艺，从城市固体废弃物中回收高价值材料，助力城市在 2040 年前实现气候中和。 - **ELECTRA**：通过电极感应加热和微波技术，推动工业高温加热炉的电气化替代改造。 - **E-FORT**：利用 AI 与自愈系统（Self-healing systems）增强关键能源基础设施的网络弹性（Cyber Resilience），降低安全风险并加速故障恢复。 - **INCIT-EV**：打造欧洲首条动态无线充电车道，为下一代电动出行提供创新充电体验。 --- ### [04:22](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=262s) 激发自下而上的创新：CIRCE 首届“AI 奥运会” 在为外部客户提供前沿数字化与绿色创新的同时，CIRCE 的管理层提出了一个自省式的核心问题：“如果我们把服务于客户的创新方法应用到自身体系中，并充分挖掘 AI 在跨学科协作中的全流程潜力，会发生什么？”为了回答“究竟该如何落地”，中心董事会发起了首届内部创新挑战赛——“AI 奥运会”（AI Olympics）。 这项为期 4 个月的内部创新赛事鼓励员工打破部门壁垒，自下而上组建跨学科团队，并为参赛者提供专属研发时间去探索兼具技术实用性与商业化潜力的 AI 方案。赛事不仅激发了全员的创造力，更诞生了一系列直击各行业实际痛点的落地应用： - **B-VI**：专为葡萄园灌溉优化设计的深度学习系统，可精准预测作物需水量并通过 WhatsApp 实时向农户推送灌溉建议，在不影响葡萄品质的前提下节水达 20%。 - **Aviguard**：利用 AI 预测未来长达 25 年的气候变化对鸟类种群迁徙和分布的影响，帮助风电场规划者规避受保护物种的高敏感区域，减少合规冲突。 - **BatLife**：电动汽车仅需接入充电桩，即可通过充电交互数据快速评估电池健康状态（State of Health, SoH）。 - **D-GEN**：实时动态评估并优化高压输电线路的输送容量上限。 - **SafeTraffic**：结合计算机视觉技术感知行人过街意图，动态调节交通信号灯配时，预计单路口每月可减少交通拥堵并减排 10 吨二氧化碳。 --- ### [07:22](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=442s) 冠军方案 La Ofertadora 的多智能体协作架构 在 AI 奥运会的所有成果中，Luis 和 Miguel Angel 主导研发的“La Ofertadora”脱颖而出夺得冠军。该系统针对科研机构编制商业与研发竞标方案时面临的流程繁琐、信息分散等问题，扮演着“AI 竞标盟友”的角色。它能帮助团队以极高的速度和严密的一致性，在短时间内检索成百上千份招标通告、提炼关键技术指标，并自动推荐量身定制的竞标报价。 La Ofertadora 依托一套完整的多智能体（Multi-Agent）系统与调度中枢（Orchestrator）实现全链路协同运作，其工作流涵盖了从公网数据获取到最终报告生成的全闭环： ```mermaid flowchart TD accTitle: La Ofertadora 多智能体协同架构图 accDescr: 展示了从公开信息抓取到最终生成竞标报告与报价推荐的多智能体工作流程。 subgraph DataIngestion ["数据获取层"] A["Web 网页爬虫与招标公告监听"] --> B["招标文件与技术附件自动下载"] end subgraph AnalysisAgent ["解析与匹配层"] B --> C["技术文档解析与核心信息抽取 Agent"] C --> D["企业内部技术能力与团队画像匹配 Agent"] C --> E["历史竞标与市场价格上下文 Agent"] end subgraph OutputLayer ["调度与交付层"] D & E --> F["调度中枢 (Orchestrator)"] F --> G["生成综合评估报告与最优价格推荐"] end ``` --- ### [08:34](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=514s) 应对大模型不确定性：演化元提示词（Evolutive Metaprompting） 第一版系统上线后，团队很快遇到了大语言模型（LLM）落地时普遍存在的随机性困境：“在 chatbot 中输入一模一样的 prompt，却往往得到不一样的回答”。在 La Ofertadora 的技术匹配环节，系统有时会对同一个任务给出不同的团队匹配得分，甚至直接更换推荐团队。这种单点的不确定性会在多智能体链条中层层放大，导致最终报告严重缺乏一致性与可靠性。 模型输出的波动性源于 LLM 底层的概率采样机制，这种伪随机性恰恰是维持模型推理能力和“创造力火花”所必需的。如果仅通过调节 temperature 等超参数强行约束输出，模型就会变得过于死板，从而丧失跨领域联想能力——这陷入了经典的“精确度（Precision）与准确度（Accuracy）的权衡”。由于多智能体之间紧密咬合，必须进行系统级的联合优化，而人工逐个测试所有提示词与参数组合根本不具可行性。 受《自然》（Nature）期刊相关研究的启发，团队提出了“演化元提示词”（Evolutive Metaprompting）机制——让 AI 智能体来优化其他智能体，实现提示词工程的自动化演化： - **双层 Agent 结构**：以“任务匹配 Agent”为基础层；在其之上设立“元提示词工程 Agent”（Meta-Prompt Engineer），负责分析基础 Agent 的执行表现并自动重写其指令与结构。 - **演化算法闭环**： 1. 初始化一个包含多种提示词与参数配置的“初始种群”； 2. 在基准测试集上评估各配置的实际表现； 3. 筛选出表现最优的提示词个体； 4. 利用大语言模型作为变异/重组算子，对其进行改写和结构变异； 5. 将新一代配置补充入种群中循环迭代，随着代际推进实现多 Agent 提示词系统的自动化全局收敛与最优调优。 ### [12:21](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=741s) 基于演化计算的 Agent 提示词多目标优化机制 在构建复杂的多 Agent 协作系统时，如何调优各个智能体的提示词与参数配置是一大核心挑战。演讲者详细介绍了他们所采用的一套基于演化计算（Evolutionary Computation）的优化流程。该流程首先从一个包含多样化配置的初始种群出发，随后在“一致性”（Consistency）与“输出质量”（Quality）两个维度上对所有配置展开评估，形成一个典型的多目标优化问题。系统筛选出表现最优的一批配置后，利用 Meta-prompting 技术模拟遗传算法中的“交叉”（Crossover）与“变异”（Mutation）过程，从而生成经过改进的新一代配置，并将其重新投入评估循环。 ```mermaid flowchart TD accTitle: 基于演化计算的 Agent 提示词优化循环 subgraph INIT["初始配置"] A["多样化配置种群<br/>(Diverse Population)"] end subgraph EVAL["多目标评估"] B["一致性评估<br/>(Consistency)"] C["质量与精度评估<br/>(Quality & Accuracy)"] end subgraph EVOL["遗传与变异"] D["筛选最优候选解"] E["Meta-prompting 交叉与变异<br/>(Crossover & Mutation)"] F["生成演化新配置与指令"] end A --> B & C B & C --> D D --> E E --> F F --> B ``` 在任务分配 Agent（Task Assignment Agent）的实测中，坐标图上的每一个点都代表一种配置方案，团队的目标是使模型状态不断向右上角的高准确率（Accuracy）与高精确度（Precision）区域靠拢。实验数据清晰地表明，仅仅经历了几代演化，进化后的提示词就全面超越了原始基线方案。这些新方案不仅在准确率上大幅提升，更展现出了极高的鲁棒性（Robustness）与稳定性（Stability）——在多个相互连接、依赖紧密的 Agent 网络中，这种稳定性是系统正常运转不可或缺的前提。 ### [13:28](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=808s) 从黑盒到可控：底层推理逻辑的质变与通用价值 通过对比演化前后的指令设计，演讲者揭示了系统性能蜕变的本质。在初始版本中，系统虽然设置了更具创造性的参数，但底层描述本身却显得十分僵化且过度依赖硬编码规则。而在经过演化算法的迭代后，表现最优的解找到了一种精妙的平衡：适度调低了创造性参数的随机度，同时搭配了更加清晰明了、兼具恰当自由裁量空间的指令体系。演讲者强调，这种方法的真正威力在于“我们不仅仅是在修饰措辞或微调参数，而是从根本上改善了系统的底层推理逻辑（underlying reasoning）”。 整个优化系统逐代调整各个 Agent 之间的协同指令，直至整个多智能体系统的整体表现达到最优。更重要的是，这套方法论不仅局限于提案生成，而是可以无缝迁移至任何基于大语言模型（LLM）的 Agent 架构中。演讲者总结了该方法的三大核心优势： - **自动化与高自适应性**：即便底层数据、具体任务或参与协作的 Agent 角色发生变更，该机制依然能够自动重新适应并持续运作。 - **结果更具鲁棒性与一致性**：大幅减少多 Agent 级联调用中的幻觉与性能波动，保证业务输出的可靠可控。 - **完全基于文本的可解释性**：所有生成与变异的解均由清晰可读的自然语言文本构成，使得人类专家能够随时进行分析与监督，彻底打破了传统模型的“黑盒”局限。 正是这种在鲁棒性、清晰度与可控性上的全面突破，才使得团队能够将一个概念性的雏形真正打磨成为可靠的业务利器；演讲者坦言，如果缺乏这一演化优化框架，“La Ofertadora”提案助手绝不可能达到今日如此稳定、出色的表现。 ### [15:07](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=907s) 业务成效落地：从商机爆发到 9 倍收入增长 算法层面的优化最终转化为业务生产力上的惊人跃升。在引入 La Ofertadora 之前的一年里，中心全年仅发掘出 86 个有潜力的研发项目商机；而在 La Ofertadora 的赋能下，团队在当年发掘出的潜在机遇直接突破了 3,000 大关。 | 业务指标 | 传统模式（基线） | La Ofertadora 赋能后潜力 | |:--- |:--- |:--- | | **商机发现量** | 86 个 | **3,000+ 个** | | **成型方案提交量** | 45 项 | **约 1,600 项** | | **最终获胜/中标项目** | 5 项 | **接近 200 项** | 在财务预测分析中，演讲者对比了常规发展路径（Business-as-usual，红线所示）与全面应用 La Ofertadora 后的收益曲线（绿线所示）。测算数据显示，在该特定业务线上，中心的年收入有望实现 **9 倍** 的爆发式增长，并且这一业务规模的扩张将直接带动创造 **60 个全新的就业岗位**，使 AI 工具的降本增效真正转化为组织扩张与实体经济的拉动力。 ### [16:03](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=963s) 组织文化变革：AI 奥运会与向内赋能的乘数效应 回顾这段探索历程，演讲者动情地表示，这不仅是一个技术项目的成功，更彻底重塑了 CIRCE 技术中心与人工智能的协作方式。在 CIRCE，AI 已经不再是孤立开展的某项特定科研任务，而是升维成为一种贯穿全局的“跨领域赋能基座”（crosscutting enabler），将顶尖人才、跨学科团队、数据资产与可持续发展目标紧密联结在一起。 这场自下而上的实践证明了一个关键论点：在组织内部开展 AI 创新并不需要动辄数以亿计的巨额资金投入，其真正的核心在于前瞻的视野、敏捷的行动力，以及拥抱实验与试错的开放文化。通过举办全员参与的“AI 奥运会”，团队领悟到了一个深刻的真理："创新不是始于实验室，也不是始于市场，而是始于自身内部（Innovation doesn't begin in the lab or in the market. It begins at home）。" 通过将这种向内自省、向内赋能的创新直接反哺到核心科研业务中——涵盖能源系统、智慧出行、数字孪生以及循环经济等诸多领域，AI 奥运会已经超越了一次单纯的黑客松或竞赛，成为推动整个机构深层次文化与组织转型的起点。当一家技术中心学会将人工智能的技术力量“向内自我应用”时，它所释放出的乘数效应，将成倍放大技术本身、人类潜能以及生态环境的长远价值。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=4s) 你每天花多少时间进行绿色思考？ — Miguel Angel Castan Lascorz > [04:22](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=262s) 如果我们把服务于客户的创新方法应用到自身体系中，并充分挖掘 AI 在跨学科协作中的全流程潜力，会发生什么？ > [08:34](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=514s) 在 chatbot 中输入一模一样的 prompt，却往往得到不一样的回答 — Miguel Angel Castan Lascorz > [13:28](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=808s) 我们不仅仅是在修饰措辞或微调参数，而是从根本上改善了系统的底层推理逻辑（underlying reasoning） — Luis Luengo > [16:03](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=963s) 创新不是始于实验室，也不是始于市场，而是始于自身内部（Innovation doesn't begin in the lab or in the market. It begins at home）。 — Miguel Angel Castan Lascorz > [16:03](https://www.youtube.com/watch?v=hgM4WSIfdE8&t=963s) 当一家技术中心学会将人工智能的技术力量“向内自我应用”时，它所释放出的乘数效应，将成倍放大技术本身、人类潜能以及生态环境的长远价值。 ## Mentioned Miguel Angel Castan Lascorz, Luis Luengo, CIRCE Technology Center, La Ofertadora, 萨拉戈萨 (Zaragoza), 皮拉尔圣母圣殿 (El Pilar), 埃布罗河 (Ebro River), REDOL, ELECTRA, E-FORT, INCIT-EV, AI 奥运会 (AI Olympics), B-VI, Aviguard, BatLife, D-GEN, SafeTraffic, Nature (期刊), WhatsApp ## Video description Miguel Angel Castan Lascorz, Project Manager in AI-based solutions for energy efficiency and sustainability, CIRCE Technology Center Luis Luengo, Project Manager, CIRCE Technology Center Abstracts What happens when a technology center decides to apply artificial intelligence... to itself? In this talk, I’ll share how CIRCE launched its first internal AI Olympics—an open innovation challenge where multidisciplinary teams developed AI-based solutions for real organizational needs. I’ll walk you through the process, the outcomes, and our most successful solution: La Ofertadora, an AI-powered assistant that’s transforming how we prepare R&D proposals. You’ll learn how AI became a practical, strategic tool within our daily workflows—without requiring massive investment. Join us to discover a replicable model of bottom-up innovation that turns AI into a catalyst for real change. [Source](https://www.youtube.com/watch?v=hgM4WSIfdE8)

Source quotes

[0:04] Miguel Angel Castan Lascorz: 你每天花多少时间进行绿色思考？ [4:22] 如果我们把服务于客户的创新方法应用到自身体系中，并充分挖掘 AI 在跨学科协作中的全流程潜力，会发生什么？ [8:34] Miguel Angel Castan Lascorz: 在 chatbot 中输入一模一样的 prompt，却往往得到不一样的回答 [13:28] Luis Luengo: 我们不仅仅是在修饰措辞或微调参数，而是从根本上改善了系统的底层推理逻辑（underlying reasoning） [16:03] Miguel Angel Castan Lascorz: 创新不是始于实验室，也不是始于市场，而是始于自身内部（Innovation doesn't begin in the lab or in the market. It begins at home）。 [16:03] 当一家技术中心学会将人工智能的技术力量“向内自我应用”时，它所释放出的乘数效应，将成倍放大技术本身、人类潜能以及生态环境的长远价值。

Source quotes

[0:04] Miguel Ángel Castán Lascorz: How much time do you spend thinking green? What happens when a technology center decides to apply artificial intelligence... to itself? [4:22] What if we apply the same innovation that we provide to our clients, to ourselves? [4:22] What if we can exploit the full power of AI into our multidisciplinary projects? [7:22] We cannot kill creativity, but we need to guide it. [15:07] vision, agility, and a culture open to experimentation and learning. [15:07] when a technology center turns its intelligence inward, it unlocks a multiplier effect a replicable model of bottom-up innovation that turns AI into a catalyst for real change.

Original transcript

[0:04] Good afternoon everybody. How much time [0:07] do you spend thinking green? [0:10] We are ML and Luis. We come from The [0:12] Technology Center in Spain and we are [0:15] both involved in energy energy [0:17] flexibility projects. We realize that [0:20] much of our time uh was spent detecting [0:23] potential opportunities, identifying [0:26] interesting sustainable projects. It [0:28] takes a lot of effort to review sources [0:30] of information and to decide if they fit [0:33] with the rest of the company's teams. So [0:36] thinking about how to increase the time [0:38] we spent on green topics is how our [0:40] story begin. We realize that we spend a [0:43] lot of time identifying green [0:45] opportunities to participate in time we [0:49] cannot devote to execute these green [0:51] projects. [0:53] Let's see some figures. Last year, we [0:56] found 22 interesting RFPs and we spent [1:01] an average of eight minutes analyzing [1:03] each of them. To tackle this inefficient [1:05] project process, we have created La [1:08] Fertadora which could be translated as [1:11] the RFPs analyzer. Since we launched it, [1:14] we were able to find 131 in just one [1:18] minute. What if we tell you that this is [1:20] an automated process that can be done by [1:23] a single click? [1:25] This is now a reality. AI helps to boost [1:28] processes and helps us to devote more [1:31] more time to green projects. Let's see [1:34] how all started. [1:36] We would like to invite you to Thraosa [1:39] Spain. [1:42] And you may be thinking Spain pa sangria [1:46] beaches. But I'm afraid I'm going to [1:49] disappoint you. Thea is not particularly [1:52] famous for any of those reasons. Our [1:55] city is special for many other. One of [1:58] them you can see it there is El Piler [2:01] and the other one is the Ero River just [2:04] beside beside it. In fact, Saragoa has [2:08] been deeply connected with water since [2:10] Asian Roman times and this connection [2:13] evolved so much that our city hosted the [2:16] World Expo in 2008. [2:18] But it's not about just about the past. [2:22] Today is exper experiencing a technology [2:26] revolution and a proof of that is the [2:29] fact that soon it will host the largest [2:32] data center in southern Europe. [2:36] TE is a clear reflection of that [2:39] process. We are a technology center with [2:43] the aim of turning environmental [2:45] challenges into innovation opportunities [2:48] and we focus on three different areas. [2:51] energy, sustainability and [2:54] digitalization. [2:55] And our goal is to help companies to [2:58] improve their competitiveness. [3:01] We thought that one of our strengths is [3:04] our diversity. At TE, we are more than [3:07] 300 professionals for a wide range of [3:10] fields. We have chemists, architects, [3:13] data scientists, software developers, [3:15] lab technicians, and much more. And we [3:18] believe that truly high impact solutions [3:21] can emerge when multi-disiplinary teams [3:23] like this work together. [3:26] Let us briefly show you what does beyond [3:29] uh artificial intelligence through four [3:31] four key international projects. Redol [3:34] applies AI, robotics and advanced [3:36] recycling to recover materials from [3:38] urban waste and to move forward to move [3:41] toward climate city neutral by 2040. [3:45] Elit supports electrification of high [3:48] temperature furnaces in industry through [3:50] electrode induction heating and [3:52] microwave technology. [3:54] Effort strength the cyber resilience of [3:57] energy infrastructure using AI and [4:00] self-healing systems to reduce risk and [4:03] speed up recovery. [4:05] And NCTV is revolutionizing EV charging [4:08] with the Europe's first dynamic wireless [4:11] charging lane. Together, these projects [4:15] show how we use technology and how to [4:17] drive in real impact across Europe. [4:22] Over the last few months, we were asking [4:24] ourselves, what if we apply the same [4:27] innovation that we provide to our [4:29] clients, to ourself, and beyond that, [4:33] what if we can exploit the full power of [4:35] AI into our multi-disiplinary projects? [4:39] And that's the starting point that [4:41] brings us here today because of course [4:44] the very next question we face was how [4:47] the hell are we going to make it [4:50] to uncover the possibilities from AI [4:53] from within the board initiated an [4:56] audocious task the AI Olympics our [5:00] international our internal initiative [5:02] centered on artificial intelligence. [5:06] The goal was bottom-up team [5:08] collaboration to develop AI powered [5:10] solutions with real market potential. [5:12] Participants were giving a dedicated [5:14] time to explore and build their ideas. [5:17] The format was a fourmonl long [5:19] competition pushing teams to go beyond [5:21] their comfort zones and learn from each [5:24] other. It became a space for creativity, [5:26] shared effort and cross-disciplinary [5:29] innovation. And beyond the outcomes, it [5:31] was an unforgettable experience for the [5:33] team. This initiative proved that giving [5:37] teams to participate to collaborate and [5:40] to embrace new opportunities AI can lead [5:42] to meaningful results. Our approach went [5:45] beyond brainstorming. We built practical [5:48] AI powered solutions that are already [5:50] transforming real processes from [5:52] different industry sectors. All the [5:55] projects have been presented to clients [5:57] and shown real market potential. Here we [6:01] can see some of our more revalent [6:04] projects. For example, we have BI which [6:07] is an app to optimize Var irrigation. It [6:11] predicts the irrigation needs using deep [6:13] learning and then it provides the [6:17] recommendation to users via WhatsApp. [6:19] The potential is 20% water savings [6:22] without affecting the quality of the [6:24] grapes. There is al also orard which [6:28] uses AI to predict how bird populations [6:31] will shift in response to climate change [6:34] projecting distribution maps up to 25 [6:37] years ahead. Ongard enables wind park [6:40] planners to avoid areas of high impact [6:43] to protected species reducing regulatory [6:46] conflicts and supporting sustainable [6:48] infrastructure development. [6:50] We also have Baltia which can evaluate [6:53] battery tests of health just like by [6:56] plugging the EV to a charger or Dragon [6:59] which can dynamically optimize the [7:00] real-time capacity of the transmission [7:03] lines. And finally we have Safeta which [7:07] modulates traffic lights according to [7:09] pedestrian crossing intents using [7:11] computer vision. It's expected to reduce [7:15] traffic congestion and 10 CO2 tons each [7:18] month per traffic like point. All these [7:22] projects were created by TE. On top of [7:25] that, we proudly stands as two of the [7:28] creators behind La Fertadora, the [7:30] champion of Tesia Olympics. As we [7:33] foreshaw in our very first slide, we [7:35] come to understand that we invest a [7:37] significant amount of time hunting for [7:39] RRPs online to bring our green projects [7:43] to life. Laerta serves as an AIdriven [7:46] ally, enabling us to craft technical [7:49] proposals with greater speed and enhance [7:52] consistency. It allows us to analyze [7:56] hundreds of proposals to detect the most [7:58] relevant bits and to suggest a price [8:01] tailored to the new opportunity. [8:03] But how can this be done? We have [8:06] created a we have created a multi- aent [8:08] system based on web scrapping and [8:12] automatic information download technical [8:14] documents analysis and information [8:16] instruction and match with the company's [8:19] technical capacities and a price [8:21] context. All this is connected through [8:24] an orchestrator that enables [8:26] communication between agents and the [8:29] result is a complete report with an [8:31] optimal price estimation [8:34] and that's all. We were happy. We [8:37] launched our first version of the system [8:40] and we thought everything was working [8:42] perfectly until we realized it wasn't. [8:45] We were facing a common issue with LLMs [8:48] that probably many of you have [8:50] experienced when working with chatbots. [8:53] Have you ever found yourselves asking [8:55] the very same question and never getting [8:57] the same result? That's what happened to [9:00] us. For example, with the agent that [9:03] matches a team to a particular task in [9:06] our first version, the agent sometimes [9:09] changed the alignment score or even [9:12] sometimes change the team. [9:14] And that's an issue. That's an issue [9:16] because this problem escalates. The [9:20] variability propagates through the [9:22] system and then the uncertainty in the [9:24] final outputs is very high. Therefore, [9:27] in the end, results lack consistency [9:31] to our solution. This was critical. But [9:34] why does this happen? Obviously, it is [9:36] not a bug. It's how these models work. [9:39] They rely on probabilistic sampling [9:41] introducing pseudo randomness in the [9:43] generation process and this is actually [9:46] necessary. It allows the model to [9:48] interpret the results. In fact, we [9:50] depend on that spark of inference to [9:53] achieve our solution. [9:55] Of course, we can reduce variability [9:57] using model parameters making outputs [9:59] more consistent. However, the cost is [10:02] high. The model becomes more robust and [10:05] loses useful connections. is the useful [10:08] problem, the useful the classic balance [10:11] between precision and accuracy. [10:14] We cannot kill creativity, but we need [10:17] to guide it. So, we require a smarter [10:20] way to reduce variability than just [10:22] through the model parameters. [10:24] Instead of relying on hard constraints, [10:26] we focused on improving the the [10:28] instructions, making prompts more [10:30] precise and align and align them with [10:33] the model parameters. [10:35] But since all agents operate together, [10:38] they need to be tuned as a whole. [10:40] Therefore, testing every possible [10:42] combination manually wasn't feasible. [10:45] That led us to a question. How can we [10:47] optimize instruction for many agents [10:49] without losing precision or creativity? [10:52] We were looking for an answer and it [10:55] came across a nature article that [10:57] sparked an idea. What if we used agents [11:01] to improve other agents? And that's the [11:04] idea behind evolutive metaprompting. A [11:07] methodology that has two elements and we [11:09] are going to take a look at them. First, [11:13] imagine an agent that assigns task to [11:16] different teams based on project [11:19] descriptions and capacities. [11:22] Then imagine an second agent whose job [11:25] is to improve the first one by rewriting [11:27] its instructions. [11:29] It plays the role of a prompt engineer. [11:32] So that based on these initial design [11:35] instructions is able to generate new [11:38] configurations for the first agent so [11:40] that it works better. [11:43] On the other hand, we have the evolutive [11:45] process to automate the improvement and [11:49] we used evol evolutionary logic to do [11:52] this. So that we began with a initial [11:55] population of different prompts and [11:56] parameters. Then we select the best one [12:00] based on their performance. And then we [12:03] generate new versions by modifying their [12:06] structure using LMS in this case [12:09] metabin. [12:10] And finally we added back to the [12:12] population and the process is repeated. [12:15] The idea is that over generations the [12:18] solutions are improved. [12:21] So our approach the process starts with [12:25] a diverse population of configurations. [12:28] Then we evaluate them on both [12:30] consistency and quality leading to a [12:33] multi-objective optimization. [12:37] Once the best configuration are [12:39] selected, they are crossed over and [12:41] mutated using meta prompting and [12:44] following the principles of evolutionary [12:46] computation. [12:48] And finally each new each new modified [12:51] solution is evaluated and the cycle [12:53] continues. [12:56] This is what we observe. We tested the [12:59] test the task assignment agent with [13:01] different evolve prompts. Each point on [13:04] the chart is a config is a configuration [13:06] and we want to reach the top right [13:08] corner high accuracy and precision. As [13:12] you can see, after a few generations, [13:15] the evolved prompts clearly outperformed [13:17] the original one. They weren't just more [13:19] accurate, they were also more robust and [13:22] more stable, which is essential when [13:24] multiple agents are interconnected. [13:28] This slide shows how the instructions [13:31] evolved. Our original version [13:35] used more creative parameters, but the [13:37] description itself was rigid and [13:39] rule-based. After the evolutionary [13:42] process, the best performing solution [13:44] found a better balance, a slightly less [13:47] creative parameters, but combined with [13:49] clear instructions that still allow for [13:52] short interpretation. [13:54] And that's the real power of this [13:56] methodology. We didn't just refine the [13:59] wording or the parameters. We improved [14:01] the underlying reasoning of the system. [14:05] And this system help us to improve the [14:07] behavior of our system. It adjusted [14:11] instructions across agents generation by [14:14] generation until the whole system [14:16] performed better. But the value goes [14:19] beyond our use case. This can be applied [14:22] to any LLM based agent system that you [14:24] may have, not just ours. First [14:29] is automatic and adaptable and it works [14:32] even when data tasks or agent change. [14:37] Second, results are more robust and [14:40] consistent. And third, it's all based on [14:44] text. So solutions can be analyzed and [14:47] supervised. And they are no longer black [14:49] boxes. Improving these core aspects, [14:52] robustness, clarity, and control. It's [14:55] what allow us to turn a promising idea [14:58] into a truly useful tool. And without [15:00] it, Laerta wouldn't perform as reliably [15:04] as it does today. [15:07] This has transformed into a tangible [15:09] phenomenon propelling our product our [15:11] productivity to new heights. The [15:13] previous year unveiled 86 promising [15:16] opportunities. Yet this year we have [15:18] uncovered more than 3,000. [15:21] From the 45 that finally made it to [15:24] submission, our contributions would have [15:26] reached around 1,600. And among the five [15:30] victories we celebrated, there lies the [15:32] potential to triumph in almost 200. [15:36] On top of this, we have calculated how [15:39] would laera affect to our annual income [15:42] in this particular line of business. The [15:44] red line represents business as usual [15:46] scenario while the green one shows our [15:50] income by using LARA. We have for seen [15:53] that we could increase nine times our [15:56] earnings and this would enhance the [15:59] creation of 60 new jobs. [16:03] This journey has transformed the way we [16:05] work with artificial intelligence at 13. [16:08] AI is no longer just another activity. [16:11] It has become a crosscutting enabler [16:13] that connects talent, teams, data, and [16:16] sustainability. And we've demonstrated [16:19] something important. Innovating with AI [16:22] doesn't require massive investment. It [16:24] requires vision, agility, and a culture [16:28] open to experimentation and learning. [16:31] Through initiative like the AI Olympics, [16:33] we've discovered something essential. [16:35] Innovation doesn't begin in the lab or [16:37] in the market. It begins at home. [16:41] And we've done it with this purpose. [16:43] Applying that innovation directly to our [16:45] projects, from energy systems to [16:48] mobility, from digital twins to circular [16:51] economy. And this AI Olympics weren't [16:53] just an experiment. They were the first [16:56] step in a deeper cultural and [16:58] organizational transformation. When a [17:01] technology center turns it its [17:03] intelligence inward, it unlocks a [17:05] multiplier effect by amplifying [17:08] technological, human, and ecological [17:11] impact. Thank you for listening. [17:13] >> Thank you.