Outdoor Equipped deploys AI agents to scale Amazon revenue 40%

Production-Grade Agents: Merging Mass Data into Real-Time Decisions With Outdoor Equipped — Ai4

Ai4

Autonomous AI agent workflows and real-time MCP data integration enable massive third-party e-commerce operations to scale multi-million dollar revenues without expanding administrative headcount.

AI-assisted summary of Production-Grade Agents: Merging Mass Data into Real-Time Decisions With Outdoor Equipped — original source. Check the original for full context.

Source published: 2026-08-19

Detailed notes

## Background Published on YouTube by Ai4 on August 19, 2026, this presentation features Eric Hankinson, VP of Technology at Outdoor Equipped. According to the page description, Hankinson walks through how he deployed AI agents and MCP servers to merge fragmented data sources across one of the largest third-party Amazon operations in the U.S., delivering real-time demand forecasting, per-transaction P&L, and automated reorders.

## TL;DR Outdoor Equipped addresses the massive data complexity of managing footwear and apparel listings on Amazon by building custom automated platforms and AI agents. The company deployed Order Desk overseen by 'Ron', an automated Hermes agent that monitors data pipelines 24/7 and auto-heals API breakdowns, alongside 'Percy', a multi-agent conversational engine powered by MCP servers and four distinct memory layers. By pre-engineering calculation scripts to save token costs and enforcing a Critique Agent refinement loop, Outdoor Equipped achieved 30% to 40% year-over-year revenue growth without scaling administrative headcount, while supplying live marketplace intelligence directly to major brand partners.

## Atmosphere & style The presentation is practical, technical, and grounded in real-world engineering rather than hyped AI promises. Hankinson speaks with the authoritative, matter-of-fact tone of an operator who has directly built, tested, and debugged these systems in a high-volume business environment. The experience feels like an insightful, transparent architectural walkthrough, highlighting cost-saving token optimization tricks and hard-won lessons alongside business successes.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=4s) Outdoor Equipped operates as a major third-party seller on Amazon, maintaining an extensive technology foundation to handle high-volume e-commerce data across over 100 brands. - [00:04](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=4s) Footwear and apparel catalog management creates high complexity, with variation child SKUs expanding single product listings into dozens or hundreds of distinct database entries. - [00:04](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=4s) Outdoor Equipped relies on three core data platforms: Precise Financial for live per-transaction P&L, Data Central for listing views and buy box trends, and MAP Watchers for monitoring price compliance. - [04:17](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=257s) The company created Order Desk to eliminate manual cross-system data assembly, pre-compiling metrics from multiple systems into automated demand and inventory reports. - [04:17](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=257s) "Ron", an AI monitoring agent built on the Hermes framework, operates 24/7 to supervise data pipelines, handle API latency, analyze failure logs, and auto-heal ingestion breaks. - [04:17](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=257s) Token efficiency is achieved by pre-engineering calculation scripts on the front end rather than relying on LLMs to read raw files and execute math dynamically. - [08:27](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=507s) "Percy" is an interactive conversational analytics tool connected to internal data sources using Model Context Protocol (MCP) servers and structured via multi-agent orchestration. - [08:27](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=507s) Percy utilizes a Coordinator Agent to break complex user queries into discrete sub-tasks, routing them to domain-specific sub-agents like Precise Financial or MAP Watchers. - [08:27](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=507s) Data outputs from sub-agents pass through an Auditor Layer for freshness validation and a Critique Agent loop for business logic evaluation and confidence scoring before being presented to users. - [08:27](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=507s) Percy incorporates four distinct memory layers: Procedural Memory for verified tool paths, Personal Memory for user roles, Company Knowledge for business rules, and Live Business Truth for real-time data access. - [14:52](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=892s) Automation and agentic workflows have allowed Outdoor Equipped to sustain 30% to 40% year-over-year revenue growth since 2012 without proportional scaling of administrative headcount. - [14:52](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=892s) Real-time data aggregation accelerates financial and inventory analytics from hours of spreadsheet manipulation down to seconds, helping preserve margins in high-volume e-commerce.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=4s) Outdoor Equipped and the Scale of E-Commerce Data

Outdoor Equipped, founded in 2012 as a startup, has expanded to become one of the largest third-party (3P) sellers on the Amazon platform. The company is family-owned and operated by Eric Hankinson, his father, and his three brothers, with Hankinson serving as VP of Technology. The business specializes in the footwear and apparel categories, acting as a fully exclusive representative for numerous major brands on Amazon. Hankinson attributes their rapid growth and strong brand partnerships to their relentless technology focus and robust data foundation, which enables them to share valuable e-commerce insights back with partner brands.

``` +------------------------------------------------------------------+ | Outdoor Equipped Architecture | +------------------------------------------------------------------+ | | | Data Foundation: | | - Precise Financial (Live post-sale fees, margins, per-SKU P&L) | | - Data Central (Real-time listing views, reviews, buy box)| | - MAP Watchers (Scrapes MAP violations & seller pricing) | | | | Data Processing & Workflow Pipelines: | | - Order Desk (Automated report aggregation & demand) | | - Ron (Hermes Agent) (24/7 pipeline monitoring & auto-healing) | | | | User Interaction & Analytics: | | - Percy (MCP Chatbot)(Multi-agent NL query engine with memory) | | | +------------------------------------------------------------------+ ```

Managing catalog data in footwear and apparel presents unique complexity compared to single-item listings (such as Thorne vitamins, which map cleanly to a single UPC row in a database). A single shoe listing with multiple colorways and sizing options can break down into 144 separate variation child SKUs. Operating at scale requires processing billions of catalog records, millions of daily marketplace and inventory observations, and representing over 100 brands—each with distinct pricing, stock levels, and margin realities.

- **Foundational Scale Metrics:** Tracks 142 million monthly product records and processes over 650,000 monthly financial event transactions. - **Core Internal Platforms:** - **Precise Financial:** Tracks live, per-transaction post-sale financial data from Amazon's API, overlaying internal costs such as fulfillment type, product dimensions, labeling, and warehouse handling fees. - **Data Central:** Provides a unified live view of listings by combining UPC details, customer review trends, seller counts, buy-box pricing, and advertising impressions. - **MAP Watchers:** A web-scraping platform that monitors Minimum Advertised Price (MAP) compliance across Amazon listings to help clean up unauthorized grey-market sellers.

---

[04:17](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=257s) Automated Demand Forecasting: Order Desk and "Ron" the Agent

To eliminate manual data assembly, Outdoor Equipped created **Order Desk**, an automated aggregation platform, overseen by **Ron**, an AI monitoring agent built on the Hermes framework. Previously, segment managers manually consolidated data from over six distinct systems into demand reports to evaluate inventory health and compute projected margins. Order Desk pre-compiles these cross-platform metrics automatically so segment teams can pull clean reports on demand.

Ron operates 24/7 to supervise data ingestion and process flows across Amazon's APIs and various brand databases. Because external API feeds and data schema frequently break or experience latency, Ron is designed to intervene, analyze failure logs, and perform corrective actions without human intervention. The data pipeline executes in four main stages:

``` [Ingest & Merge] ---> [Data Validation] ---> [Report Formatting] ---> [Reorder Engine] (Multi-source feeds) (Source of Truth) (Color-coded UI) (Automated Recs) ```

1. **Ingest & Merge:** Pulls raw data from all active sources and consolidates them into a unified order-per-SKU record. 2. **Validation:** Verifies data freshness and integrity to ensure it adheres to the company's single source of truth. 3. **Report Generation:** Formats demand sheets with visual color coding and custom filters so segment managers can rapidly interpret trends. 4. **Reorder Suggestion:** Generates automated SKU-level reorder recommendations per brand based on historical velocity and inventory movement.

Hankinson highlights key architectural design choices regarding AI agent token efficiency: rather than tasking an LLM with reading raw files and dynamically reasoning through calculations—which "ate tokens and wasn't as accurate"—they pre-engineered calculation scripts on the front end. AI tokens are spent primarily during script generation and tool connection setup, leaving Ron to run idle and consume tokens only when intervention is required.

---

[08:27](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=507s) Interactive Multi-Agent Analytics: Meet "Percy"

To replace manual SQL/query-builder filters and Excel exports, Outdoor Equipped introduced **Percy**, an integrated conversational AI chatbot connected to disparate internal data sources via Model Context Protocol (MCP) servers and structured through a multi-agent orchestration workflow.

``` +-----------------------+ | User Prompt | +-----------+-----------+ | v +-----------------------+ | Coordinator Agent | | (Intent/Task Parsing) | +-----------+-----------+ | +------------------+------------------+ | | | v v v +---------------+ +---------------+ +---------------+ | Sub-Agent: | | Sub-Agent: | | Sub-Agent: | | Precise Fin. | | Data Central | | MAP Watchers | +-------+-------+ +-------+-------+ +-------+-------+ | | | +------------------+------------------+ | v +-----------------------+ | Auditor Layer | | (Source Validation) | +-----------+-----------+ | v +-----------------------+ | Critique Agent | | (Business Logic Eval) | +-----------+-----------+ | [Pass Score] / \ [Low Score -> Retry Loop] v +-----------------------+ | Final User Answer | +-----------------------+ ```

When a user submits a query (for example, analyzing monthly return rates for Asics footwear over the past year), the query passes through several multi-agent handling layers:

- **Message Parsing & Coordinator Agent:** Breaks complex queries into discrete intent tasks, execution rules, and evidence requirements before dispatching tasks to domain-specific sub-agents (e.g., Precise Financial, Data Central, or MAP Watchers). - **Sub-Agent Execution & Auditor Layer:** Sub-agents execute data retrievals against their assigned MCP server endpoints. An auditor layer validates that data sources are fresh, authentic, and complete. - **Generator & Critique Agent Loop:** Before returning a response, a Critique Agent evaluates the result against internal business logic and assigns a confidence score. If the output fails business rules or lacks actionable depth, it enters a re-try refinement loop. Once a high confidence score is achieved, the answer is released to the user.

Percy's reasoning accuracy is reinforced by four dedicated memory layers: - **Procedural Memory:** Remembers verified tool paths for frequent query types (like return-rate calculations), bypassing redundant planning steps to accelerate query response times. - **Personal Memory:** Tailors responses to the specific functional role of the prompt originator (e.g., logistics, brand management, listings, or ad performance teams). - **Company Knowledge:** Embeds domain context around brand contracts, marketplace nuances, seasonality patterns, pricing dynamics, and catalog dispute resolution workflows. - **Live Business Truth:** Guarantees sub-agents always query fresh, real-time platform data rather than stale cached states.

---

[14:52](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=892s) Business Impact and Strategic Outcomes

Implementing production-grade agentic workflows and automated data aggregation has driven notable operational efficiencies across Outdoor Equipped's business:

- **Headcount Scale Efficiency:** Supported consistent year-over-year revenue growth of 30% to 40% since 2012 without requiring proportional increases in administrative staff headcount, enabling talent redeployment to higher-value roles. - **Real-Time Decision Making:** Accelerated financial and inventory analysis from hours of spreadsheet manipulation down to seconds—critical in Amazon's thin-margin environment where rapid repricing, fee audit corrections, and seasonal velocity shifts impact overall profitability. - **Enhanced Brand Partnerships:** Shared granular, real-time marketplace insights directly with major vendor partners—including Asics, Thoroughgood, and The North Face—helping them optimize inventory forecasting, product design, and off-Amazon marketing strategies.

## Notable quotes > In high-volume operations like Amazon retail, the gap between data and decision is where margin quietly disappears — and traditional BI was never built to close it. — Eric Hankinson

> What if your business could make every decision the moment the data arrived, instead of waiting on a report? — Eric Hankinson

## Mentioned Outdoor Equipped, Eric Hankinson, Amazon, Precise Financial, Data Central, MAP Watchers, Order Desk, Ron, Hermes Agent, Percy, Model Context Protocol, Asics, Thoroughgood, The North Face, Thorne

## Video description Eric Hankinson - VP of Technology, Outdoor Equipped

What if your business could make every decision the moment the data arrived, instead of waiting on a report? In high-volume operations like Amazon retail, the gap between data and decision is where margin quietly disappears — and traditional BI was never built to close it. I will walk through how I deployed AI agents and MCP servers to merge fragmented data sources across one of the largest third-party Amazon operations in the U.S., delivering per-transaction P&L, demand forecasting, and reorder decisions in real time. Attendees will leave with a practical understanding of how to architect agentic workflows for production, the pitfalls I hit along the way, and the design choices that separate demo-ware from systems that actually run a business. If you're responsible for decisions inside a data-heavy operation, this talk will show you what's already possible — and what's coming next.

[Source](https://www.youtube.com/watch?v=z2IopjXwm8Y)

中文简报

Outdoor Equipped部署AI智能体，推动亚马逊营收增长40% 自主AI智能体工作流与实时MCP数据集成助力大型第三方电商业务实现数百万美元的营收扩张，且无需增加行政人员编制。 Outdoor Equipped通过构建定制自动化平台与AI智能体，化解了在亚马逊上管理鞋服商品信息的海量数据复杂性难题。该公司部署了由“Ron”负责监管的Order Desk系统，后者是一个基于Hermes框架的自动化智能体，可全天候监控数据管道并自动修复API故障；同时还部署了由MCP服务器与四个独立记忆层驱动的多智能体对话引擎“Percy”。通过预先设计计算脚本以节约Token成本，并执行Critique Agent优化闭环，Outdoor Equipped在未扩增行政人员编制的前提下实现了30%至40%的年营收同比增长，同时直接向主要品牌合作伙伴提供实时市场情报。 Outdoor Equipped是亚马逊上的大型第三方卖家，依托深厚的技术基础处理覆盖100多个品牌的高体量电商数据。 (0:04) 鞋服类目管理极其复杂，变体子SKU会将单一商品页面拓展为数十乃至数百个独立的数据库条目。 (0:04) Outdoor Equipped依赖三大核心数据平台：用于实时统计单笔交易损益的Precise Financial、用于追踪商品浏览量与Buy Box趋势的Data Central，以及用于监控最低广告价格合规性的MAP Watchers。 (0:04) 该公司创建了Order Desk以消除人工跨系统数据整合，将多系统指标预先编译生成自动化的需求与库存报告。 (4:17) 基于Hermes框架构建的AI监控智能体“Ron”全天候运行，负责监督数据管道、处理API延迟、分析故障日志并自动修复数据摄入中断问题。 (4:17)

Reading summary

Outdoor Equipped VP of Technology Eric Hankinson describes how the high-volume Amazon seller combined a proprietary data foundation with specialized autonomous agents to automate demand forecasting and operational decision-making. Catalog Complexity and Core Data Platforms (0:04) Operating as a third-party seller for over 100 footwear and apparel brands introduces severe data complexity. Unlike simple single-UPC listings, apparel products involve extensive variations; a single shoe listing can span 144 separate child SKUs across sizes and colorways. To manage this volume, the company processes 142 million monthly product records and more than 650,000 monthly financial transactions across several custom internal tools. These platforms include Precise Financial, which pulls transaction-level post-sale figures from Amazon's API and applies internal metrics such as dimensions, labeling, and warehouse fees; Data Central, which consolidates buy-box pricing, customer reviews, seller counts, and advertising impressions; and MAP Watchers, a scraping engine designed to detect Minimum Advertised Price infractions across the marketplace. Pipeline Automation and Token-Efficient Maintenance (4:17) To replace the manual consolidation of reports across six external platforms, Outdoor Equipped established Order Desk, an automated aggregation pipeline overseen by an autonomous Hermes-based agent named Ron. The system processes records through a four-stage workflow: multi-source ingestion and merging, validation against the company's single source of truth, color-coded report formatting, and automated brand-specific reorder recommendations based on historical velocity. To optimize operating costs, Hankinson notes that the team intentionally avoided having large language models parse raw files or calculate figures dynamically, finding that it drained tokens while degrading calculation accuracy. Instead, calculations are handled by pre-engineered front-end scripts. AI tokens are reserved for generating scripts and configuring tool connections, allowing the agent to remain idle until external API schemas break or latency requires automated log analysis and pipeline self-healing. Multi-Agent Query Architecture and Memory Systems (8:27) For internal queries, the company deployed Percy, an interactive chatbot connected to internal databases through Model Context Protocol (MCP) servers. Incoming prompts are parsed by a Coordinator Agent, which isolates execution rules and dispatches requests to domain sub-agents linked to Precise Financial, Data Central, or MAP Watchers. The data passes through an Auditor Layer to confirm authenticity and freshness, followed by a Critique Agent that evaluates output against internal business logic and triggers refinement loops if confidence thresholds are not met. Percy's accuracy across departments is maintained through four dedicated memory structures. Procedural memory stores pre-validated tool paths for recurring tasks like return-rate calculations to reduce latency; personal memory adapts responses to the user's operational role; company knowledge incorporates partner contracts, marketplace nuances, and dispute procedures; and live business truth prevents stale caching by mandating real-time database queries. Operational Efficiency and Brand Collaboration Gains (14:52) Hankinson reports that automated agent workflows have enabled Outdoor Equipped to sustain 30% to 40% year-over-year revenue growth since 2012 without requiring proportional increases in administrative staff headcount. Routine inventory and margin analyses that previously took hours of spreadsheet compilation are delivered in seconds, providing the responsiveness needed to protect margins against shifting fees and repricing dynamics. Additionally, granular real-time marketplace metrics are shared directly with exclusive vendor partners such as Asics, Thoroughgood, and The North Face. These data feeds assist brand partners with production planning, inventory forecasting, and external marketing strategies outside the Amazon ecosystem. Outdoor Equipped illustrates how production agents succeed when paired with pre-engineered deterministic scripts and layered memory architectures, though maintaining these systems requires continuous operational oversight as external platform APIs and marketplace data schemas shift.

精炼阅读摘要

Outdoor Equipped技术副总裁Eric Hankinson阐述了该企业如何针对鞋服电商复杂的变体数据与庞大的交易量，搭建集成底层数据库、具备自愈能力的流水线监控代理Ron以及基于MCP协议的多代理查询系统Percy，将数亿级原始数据转化为支撑秒级运营决策的实时能力。多变体商品架构与核心数据基座 (0:04) 鞋类与服装品类的电商目录管理具有极高的数据复杂度。不同于Thorne维生素等单UPC商品，一双鞋因颜色和尺码差异即可衍生出多达144个子变体SKU。作为代表100多个品牌的亚马逊第三方独家合作卖家，Outdoor Equipped每月需追踪1.42亿条商品记录并处理超过65万笔交易事件。针对庞大的体量与定价差异，公司构建了三大底层自研系统来支撑业务运营。具体而言，Precise Financial通过对接亚马逊API实时追踪每笔售出交易费用，并叠加仓储包装、尺寸测量与标签处理等内部实际履约成本，测算出单SKU维度的净损益；Data Central汇集了商品UPC详情、买家评价趋势、卖家数量、购物车（Buy Box）占有及广告曝光量，提供动态总览视图；MAP Watchers则借助网络爬虫持续监测最低广告价格（MAP）的执行情况，协助品牌清理未经授权的灰色市场销售行为。自动化需求流水线与Ron代理容错 (4:17) 为终结跨6套以上系统手动制作报表的低效流程，公司开发了自动化汇总平台Order Desk，贯通了多源数据摄取合并、真实源验证、可视化需求报表生成，以及基于动销速度的品牌SKU自动补货建议四个环节。运行于Hermes框架上的AI代理“Ron”则执行7×24小时全天候巡检，监控亚马逊API及品牌数据库的数据流。当遇到外部API模式变更或延迟报错时，Ron能够自主分析故障日志并执行修正，实现流水线的无人化自愈修复。Hankinson指出，在Token成本与计算效率控制上，他们并未让大语言模型直接读取原始文件去进行复杂的数学计算，而是将计算逻辑固化在前端预编写的脚本中；AI Token仅在初始生成脚本、对接工具和发生异常干预时消耗，使Ron在平时处于极低开销的待命状态。 Percy多代理协同与四层记忆系统 (8:27) 为取代传统SQL查询与Excel导出，公司推出了自然语言对话机器人Percy。该系统借助模型上下文协议（MCP）服务器连接各异构数据库，并通过多代理架构处理业务提问。用户的查询先由协调代理（Coordinator Agent）进行意图拆解与规则界定，随后分派给Precise Financial、Data Central或MAP Watchers等专属子代理；抓取到的数据须经过审计层验证真实性与时效性，再交由审查代理（Critique Agent）比对业务逻辑评分，若未达到可执行深度则触发重新优化循环。 Percy的推理可靠性依托于四层记忆机制：程序记忆记录高频查询（如退货率计算）的已验证工具调用路径以跳过冗余规划；个人记忆依据提问者所属部门（如物流、品牌管理、运营或广告投放）定制回答视角；公司知识沉淀了品牌合同细则、季节性特征与目录纠纷流程；实时商业真值则强制子代理每次均读取即时数据，防止出现基于过期缓存的幻觉决策。规模化人效增长与品牌实时洞察赋能 (14:52) 生产级智能代理工作流的落地为企业带来了显著的运营红利。据Hankinson介绍，自2012年创立以来，公司实现了30%至40%的年复合营收增长，且无需同比例增加行政人员编制，实现了人员向更高价值岗位的优化配置。在利润微薄的亚马逊电商生态中，原本耗费数小时的手工表格分析被缩短至数秒完成，确保了针对动态调价、平台费用核查与季节性动销波动的实时应对。此外，公司直接向包括Asics、Thoroughgood及The North Face在内的品牌供应商开放精细化的实时市场洞察数据，反向赋能品牌方的整体库存预测、商品设计研发以及亚马逊站外的全渠道营销战略布局。该案例展示了针对特定行业痛点定制自动化代理及前置脚本计算的工程效益，但系统高度依赖于亚马逊底层API的接口开放度与数据流稳定性；如何在更多外部数据接口存在突发波动的环境下持续保障代理的自治健壮性，仍是此类自动化系统长期演进的核心课题。

Detailed notes — zh

## Background 本内容由 Ai4 频道于 2026 年 8 月 19 日在 YouTube 平台发布，演讲者为 Outdoor Equipped 技术副总裁 Eric Hankinson。视频主要介绍了其团队如何部署生产级 AI 智能体与 MCP 服务器整合海量分散数据，在全美规模领先的亚马逊第三方鞋服零售业务中实现单笔交易 P&L 结算、需求预测与秒级实时补货决策。 ## TL;DR Outdoor Equipped 针对鞋服电商中商品变体庞大、退货率高且利润微薄的痛点，构建了 Precise Financial、Data Central 与 MAP Watchers 三大核心数据底座。团队开发了 24/7 监控与自愈智能体 Ron 以保障 ETL 管道的单一事实来源，并打造了基于 MCP 协议与多智能体架构的对话式 AI 系统 Percy。通过引入评估循环、程序性记忆、个人记忆及企业领域知识分层，系统能够以低 Token 成本输出高置信度的业务决策，在保持年均 30%–40% 业务增速的同时避免了人员规模的线性扩张，并将实时运营洞察反哺给 ASICS、The North Face 等合作品牌。 ## Atmosphere & style 整场演讲风格扎实硬核、务实接地气，充溢着资深一线架构师直面业务痛点的工程师底色。演讲者摒弃了空洞的概念炒作，以家族企业在亚马逊平台搏杀的真实业务场景为切入点，逻辑严密地拆解了从底层海量数据流到多智能体自愈、评估与记忆系统的演进脉络。现场节奏紧凑而充满洞见，生动展现了生产级 AI 如何在严苛的微利零售环境中真正转化为企业的商业护城河。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=4s) Outdoor Equipped 深耕亚马逊鞋服品类，单个商品 Listing 可细分出多达 144 种尺码与颜色变体，导致数据库和分析处理的数据维度呈几何级数激增。 - [01:46](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=106s) 团队构建了三大核心基础设施：精准还原售后毛利的 Precise Financial、监控 Listing 全貌的 Data Central 以及全天候抓取竞品价格防窜货的 MAP Watchers。 - [01:46](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=106s) 数据中台每月需处理超过 1.42 亿条产品动态及 65 万笔售后交易，支撑百余个品牌各自独立的利润与定价规则的毫秒级追踪。 - [04:17](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=257s) 针对运营人员需跨 6 个以上数据源预测需求的痛点，集中化平台 Order Desk 引入全天候智能体 Ron，自动化监控流转管道并在数据异常时自主修复，确立单一事实来源。 - [04:17](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=257s) 工程架构避开将大量原始数据直接抛给 LLM 的高 Token 消耗误区，通过确定性脚本与工具完成前置数据清洗，仅在管道断流和异常干预时触发大模型调用。 - [08:27](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=507s) 自研对话式系统 Percy 依托模型上下文协议（MCP）整合多智能体网络，由协调智能体拆解意图并路由至对接各业务系统的子智能体获取精准数据。 - [08:27](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=507s) 引入审计层验证数据时效，并通过生成器与评估智能体闭环（Generator-Critique Loop）根据预设业务逻辑严格打分，避免模型生成脱离业务实际的冷冰结论。 - [08:27](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=507s) 采用程序性记忆（Procedural Memory）沉淀经过验证的成功查询链路，后续相同业务场景可直接复用成熟路径，大幅降低推理延迟与推理成本。 - [13:12](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=792s) 构建多层记忆体系，涵盖物流与广告等不同角色的个人记忆，结合岗位长远目标定制化生成符合各自业务视角的差异化回答。 - [13:12](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=792s) 将平台政策博弈、品牌合作特征及鞋服季节性动销规律沉淀为企业领域知识注入智能体，实现对宏观业绩波动的多维度即时深度归因。 - [14:47](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=887s) 在公司年均实现 30%–40% 复合增长的同时，智能体接管繁杂自动化流程打破了人员编制的线性膨胀，释放人力聚焦高价值战略业务。 - [14:47](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=887s) 通过秒级动态调价与违规政策管控，化解鞋服品类极高退货率与微利属性带来的利润挤压，并敏捷调拨自有仓与 FBA 库存应对季节动销节奏。 - [14:47](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=887s) 将一手实时动销数据反哺给 ASICS、The North Face 等核心合作品牌，直接赋能品牌下一代产品研发、营销及产能预测，构筑深度绑定的商业护城河。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=4s) Outdoor Equipped 背景与鞋服品类的复杂数据挑战 Outdoor Equipped 成立于 2012 年，最初只是一家初创公司，如今已成长为亚马逊平台上规模最大的第三方卖家（3P Seller）之一。公司采用家族式经营模式，由演讲者 Eric Hankinson、他的父亲以及三位兄弟共同管理，Eric 主要负责技术研发。公司专注于鞋类和服装（Footwear and Apparel）品类，获得了众多知名品牌的独家代理权，在亚马逊平台上全方位代表这些品牌运营。支撑其高速增长和品牌深度合作的核心，正是公司自主构建的技术底座与数据共享体系。在电商零售中，不同品类的数据库架构复杂度有着天壤之别。以常规标品（如 Thorne 维生素）为例，一个商品页面仅对应一个通用产品代码（UPC），数据库只需记录一行数据及流入的信息即可。然而，鞋类和服装品类极其繁琐，同一款商品可能衍生出繁多的颜色和尺码组合。在展示的一个商品页面案例中，单个 Listing 下细分出了多达 144 种变体选项。这意味着无论是构建底层数据库、进行需求预测，还是抓取市场情报，数据处理的体量与维度都会呈几何级数增长。 ### [01:46](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=106s) 海量数据规模与三大核心基础设施平台由于商品变体极其庞大，公司的日常运营面临着极高的数据吞吐压力。为了支撑实时决策并摸清真实的利润与动销情况，团队构建了三大核心底层系统，打通从销售端到供应链的全链路数据： ```mermaid flowchart TD accTitle: 数据底座三大核心系统架构 subgraph DataIngestion ["底层数据平台"] PF["Precise Financial 实时售后财务追踪"] DC["Data Central Listing 360° 实时大盘"] MW["MAP Watchers 价格监控与防窜货爬虫"] end subgraph Processing ["业务赋能"] PF --> Cost["单笔交易级 P&L 结算"] DC --> Market["转化率/竞品/广告分析"] MW --> Enforce["品牌最低限价 MAP 维权"] end ``` - **Precise Financial（精准财务平台）**：亚马逊的交易数据如同一个“动态生命体”，在订单提交、信用卡扣款到最终结算的不同时间节点会不断补充消费者信息。该系统通过实时调用亚马逊 API 获取单笔交易级数据，并叠加公司内部的履约方式、商品尺寸规格、品牌分摊仓储费、贴标重包费、广告投放成本以及商品销售成本（COGS），还原真实的单笔交易售后利润。 - **Data Central（数据中心）**：提供 Listing 的 360 度实时全貌，汇聚 UPC 属性、评论数量、在售卖家数、跟卖价格、广告曝光量（Impressions）等多维度指标。 - **MAP Watchers（最低限价监控）**：在高端品牌运营中，维持最低转售价（MAP）至关重要。亚马逊平台有时如同“狂野西部（Wild West）”，许多夫妻小店（Mom-and-pop shops）或未授权卖家会隐匿身份倾销库存。MAP Watchers 通过网络爬虫全天候抓取竞品价格波动，量化价格破价对整体销量的冲击，协助品牌方清理违规卖家。 - **核心数据资产规模**： - 沉淀数十亿条商品目录记录（Catalog Records）； - 每日新增数百万条市场与库存监控观测值； - 深度代理追踪 100 多个品牌，分别对应独立的价格、库存与利润规则； - 每月追踪 1.42 亿条产品动态记录； - 每月处理超过 65 万笔售后财务事件交易。 ### [04:17](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=257s) Order Desk 与 24/7 监控智能体 Ron 对于类目运营经理（Segment Managers）而言，进行科学的需求预测和补货极其繁琐，他们此前需要从至少 6 个分散的数据源手动汇总数据。为此，团队开发了集中化报表平台 **Order Desk**，并部署了名为 **Ron** 的全天候 AI 智能体来自动化接管整个 ETL 流程。智能体 Ron 基于 Hermes 架构构建，专门负责监控数据集成与流转管道。由于系统需要持续从亚马逊 API 以及各大品牌方各自独立的数据库中提取数据，接口异常和数据中断屡见不鲜。Ron 能够在后台实时侦测故障节点并自动执行修复，确保最终生成的每一份需求报表包含 Listing 评价、曝光量、总销量、当前库存、品牌方可用库存（ATS）以及预估利润率等关键要素，并形成唯一的“单一事实来源（Single Source of Truth）”。 ```mermaid flowchart LR accTitle: Order Desk 与智能体 Ron 的自动化补货流 subgraph Ingestion ["数据摄取与对齐"] A["6+ 异构数据源"] --> B["按 SKU 归并整合"] end subgraph Governance ["治理与监控"] B --> C["数据校验 & 真实源确认"] RonAgent["智能体 Ron (24/7 监控与自愈)"] -.->|监测与修复| B RonAgent -.->|异常拦截| C end subgraph Output ["决策输出"] C --> D["需求数据报告"] D --> E["自动生成补货建议"] end ``` 在工程架构设计上，团队避开了一个常见误区。很多人认为只要把数据直接丢给大模型并提示“随便处理这些数据并给出好建议”即可，但这种做法极其消耗 Token，且输出结果往往不够精确。团队选择将 Token 消耗前置到工具与连接器的开发阶段，大量数据清洗与聚合计算全部由前端确定性脚本和公式完成；智能体 Ron 仅作为监管与干预者存在，只在数据断流或需要介入修正时才触发模型消耗 Token。最终，运营人员收到的报表已自带针对各品牌的建议补货单，经理只需审核智能体基于历史销速与库存周转给出的推荐理由即可。 ### [08:27](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=507s) Percy：基于 MCP 的多智能体查询与记忆架构在 Precise Financial 早期，运营经理若想针对特定 SKU 或品牌进行深度剖析，必须手动在查询构建器（Query Builder）中配置规则群组、导出 Excel 并自行套用复杂公式。为了消除这一分析鸿沟，团队打造了名为 **Percy** 的内置对话式 AI 系统。Percy 通过模型上下文协议（MCP）服务器连接所有分散的业务数据库，并驱动一套多智能体工作流。 ```mermaid flowchart TD accTitle: Percy 多智能体协同查询与评估流程 User["用户输入自然语言查询 (例如：ASICS 跑鞋月度退货率对比)"] --> Coord["协调智能体 (Coordinator Agent) 拆解意图/子任务/证据需求"] subgraph SubAgents ["MCP 子智能体层"] Coord --> SA1["Precise Financial 智能体"] Coord --> SA2["Data Central 智能体"] Coord --> SA3["MAP Watchers 智能体"] end SubAgents --> Audit["审计验证层 (Auditor Layer) 验证数据源真实性与时效"] Audit --> GenLoop["生成器 (Generator)"] subgraph CritiqueWorkflow ["业务逻辑质量评估环"] GenLoop --> Critique["评估智能体 (Critique Agent) 注入业务逻辑打分"] Critique -- "低分：重新计算/补充证据" --> GenLoop end Critique -- "高分：交付置信结论" --> Output["交付最终洞见与建议"] subgraph Memory ["记忆层"] Output -.-> PM["程序性记忆 (Procedural Memory) 沉淀为经过验证的固定执行路径"] PM -.->|跳过重复推理| Coord end ``` 当团队成员向 Percy 提出宽泛的分析需求（例如：“查询 ASICS 跑鞋本月的退货率，并与过去一年的各月份进行对比”）时，系统的各个层级会协同响应： - **协调智能体（Coordinator Agent）**：负责意图拆解，将自然语言转化为具体任务清单和证据需求（如分别提取各月总销售量与总退货量、套用退货率计算公式、按时间序列进行横向比对）。 - **子智能体路由（Sub-Agent Routing）**：协调智能体根据证据需求，将请求分发给对接 Precise Financial、Data Central 或 MAP Watchers 的不同子智能体，由其通过 MCP 调取精准数据。 - **审计验证层（Auditor Layer）**：负责校验抓取数据的时效性与数据源合法性，避免模型基于过期数据进行推演。 - **评估反馈循环（Generator-Critique Loop）**：系统不仅提供冷冰冰的数字，更追求输出具备业务洞察的建议。评估智能体（Critique Agent）会依据预设的业务逻辑对初版答案进行严格打分，评分不达标则打回循环重新生成，只有高分答案才会被交付给终端用户。 - **程序性记忆（Procedural Memory）**：这是整个多智能体架构提效的关键。当多次出现同类查询（如按月度计算品牌退货率）时，系统会将其固化为一条“已验证路径（Verified Path）”。后续再遇到相似提问时，智能体无需再从零拆解推理，而是直接调用成熟路径，大幅降低了推理延迟与 Token 成本。 ### [13:12](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=792s) 打造多层记忆系统：从用户画像到企业领域知识在构建生产级智能体时，数据获取路径的记忆与优化至关重要。当智能体探索并验证了一条获取特定数据的成功路径后，系统会记住这种“经过验证的有效方法”，并在后续调用中精简流程，大幅提升响应速度。在此基础上，智能体“Percy”引入了精细化的分层记忆机制，首先是“个人记忆”（Personal Memory）。在 Outdoor Equipped 内部，物流团队、板块业务经理（Segment Managers）、商品上架部门（Listings Department）以及广告投放部门都会向 Percy 提交各种业务查询。通过在单个用户层面沉淀记忆，Percy 能够清晰识别提问者的岗位角色及其长期的业务目标与诉求偏好，从而针对不同员工“量身定制”符合其业务视角的回复。 ```mermaid flowchart TD accTitle: Percy 智能体多层记忆与决策架构 subgraph MemoryLayers["Percy 记忆与知识分层"] A["路径记忆 (记住已验证的数据抓取路径)"] B["个人记忆 (识别提问角色与业务目标)"] C["企业领域知识 (品牌关系、季节周期与平台博弈)"] D["实时业务真相 (保持各系统数据毫秒级同步)"] end subgraph DecisionEngine["智能推理与决策输出"] E["跨部门复杂业务提问 ('某品牌销量为何环比波动？')"] F["多维度归因分析引擎"] G["定制化即时业务决策"] end MemoryLayers --> F E --> F F --> G ``` 更深的一层则是“企业领域知识”（Company Knowledge）。电商代运营与经销业务极为复杂，每个合作品牌都有着独特的业务关系，同时在亚马逊平台上还要应对五花八门的现实挑战——比如商品 Listing 突发故障、与亚马逊官方规则的摩擦博弈、跟卖与定价波动、市场大盘变动以及服装鞋履固有的季节性周转速度（Seasonal Velocity）。通过将这些深度的企业背景与行业认知预先注入智能体的记忆层，员工在提问时只需抛出宏观问题，例如“能否告诉我这个品牌当下的销售额为什么会出现上升或下滑，或者与上个月相比表现如何？”系统便能综合考量复杂的市场环境、品牌特征、已知阻碍以及最新的“实时业务真相”（Live Business Truth），在底层数据随时刷新同步的前提下，迅速完成深度归因并给出精准回答。 - **个人记忆赋能跨职能协作**：涵盖物流、品类管理、Listing 运营及广告团队，智能体根据用户过往交互习惯和岗位目标差异化输出分析深度与业务侧重点。 - **融合复杂业务上下文**：将平台政策博弈、价格变动机制、淡旺季动销规律等非结构化企业认知转化为智能体的推理依据。 - **实时业务基准保障**：强调业务底层数据的实时更新与无缝同步，杜绝基于过时信息做判断的“幻觉”。 --- ### [14:47](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=887s) 业务收益与品牌赋能：在微利与高退货率中实现秒级决策部署生产级智能体为企业组织带来了立竿见影的结构性红利。自 2012 年以来，Outdoor Equipped 始终保持着每年 30% 到 40% 的复合增长率，但借助智能体对高频分析与自动化流程的接管，企业成功将现有员工重新分配到更具战略价值的业务板块，实现了在业务体量连年高速扩张的同时无需按比例扩招团队。在运营端，亚马逊平台的零售生态本质上是一个“微利业务”（Micro-margin business），在公司深耕的服装与鞋类类目中尤为如此，不仅利润空间受到严格挤压，还伴随着极高的退货率挑战。智能体实现的“秒级决策”，使团队能够以极高的频率监控全局数据，迅速通过调整动态定价或对违规品牌推进政策落地，最大限度锁定并扩大整体利润率。在库存周转方面，服装行业的季节性特征要求供应链必须具备“雷厉风行的敏捷反应”，无论是调配自身仓储物流还是调整亚马逊 FBA 仓库中的在售状态，系统都能基于实时流速迅速做出库存调拨决策。更深远的影响在于品牌合作关系的升维：通过将庞大细致的实时运营数据反哺给合作伙伴，ASICS、Thorogood 以及 The North Face 等关键合作品牌能够直接将这些一手市场洞察接入自身的业务系统，为下一代产品的设计取向、市场营销策略以及生产库存预测提供高价值的数据支撑。在亚马逊作为全球最大电商交易平台的背景下，这种将海量分散数据转化为即时决策力、并深度绑定头部品牌全域电商战略的能力，已然成为企业最坚固的商业护城河。 - **人力杠杆效应显著**：在年均保持 30%–40% 业务增速的背景下，避免了人员编制的线性膨胀，将人力资源重新聚焦于高价值业务节点。 - **秒级应对微利挑战**：针对鞋服品类的高退货率与微利属性，依靠实时洞察动态调节前台定价与品牌管控规则，守护单品毛利空间。 - **敏捷库存流动控制**：配合季节性动销节奏，即时联动自有仓与亚马逊 FBA 仓储网络，快速切换库存激活状态。 - **重塑品牌战略协同**：为 ASICS、Thorogood、The North Face 等品牌输出实时消费与运营数据，直接参与品牌端的设计研发、营销规划与供应链预测。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=4s) 一个商品页面仅对应一个通用产品代码（UPC），数据库只需记录一行数据及流入的信息即可。然而，鞋类和服装品类极其繁琐，同一款商品可能衍生出繁多的颜色和尺码组合。 — Eric Hankinson > [01:46](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=106s) 亚马逊的交易数据如同一个“动态生命体”，在订单提交、信用卡扣款到最终结算的不同时间节点会不断补充消费者信息。 — Eric Hankinson > [01:46](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=106s) 亚马逊平台有时如同“狂野西部（Wild West）”，许多夫妻小店（Mom-and-pop shops）或未授权卖家会隐匿身份倾销库存。 — Eric Hankinson > [04:17](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=257s) 很多人认为只要把数据直接丢给大模型并提示“随便处理这些数据并给出好建议”即可，但这种做法极其消耗 Token，且输出结果往往不够精确。 — Eric Hankinson > [04:17](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=257s) 团队选择将 Token 消耗前置到工具与连接器的开发阶段，大量数据清洗与聚合计算全部由前端确定性脚本和公式完成；智能体 Ron 仅作为监管与干预者存在，只在数据断流或需要介入修正时才触发模型消耗 Token。 — Eric Hankinson > [08:27](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=507s) 当多次出现同类查询（如按月度计算品牌退货率）时，系统会将其固化为一条“已验证路径（Verified Path）”。后续再遇到相似提问时，智能体无需再从零拆解推理，而是直接调用成熟路径，大幅降低了推理延迟与 Token 成本。 — Eric Hankinson > [13:12](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=792s) 通过在单个用户层面沉淀记忆，Percy 能够清晰识别提问者的岗位角色及其长期的业务目标与诉求偏好，从而针对不同员工“量身定制”符合其业务视角的回复。 — Eric Hankinson > [13:12](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=792s) 员工在提问时只需抛出宏观问题……系统便能综合考量复杂的市场环境、品牌特征、已知阻碍以及最新的“实时业务真相”（Live Business Truth），在底层数据随时刷新同步的前提下，迅速完成深度归因并给出精准回答。 — Eric Hankinson > [14:47](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=887s) 自 2012 年以来，Outdoor Equipped 始终保持着每年 30% 到 40% 的复合增长率，但借助智能体对高频分析与自动化流程的接管，企业成功将现有员工重新分配到更具战略价值的业务板块，实现了在业务体量连年高速扩张的同时无需按比例扩招团队。 — Eric Hankinson > [14:47](https://www.youtube.com/watch?v=z2IopjXwm8Y&t=887s) 亚马逊平台的零售生态本质上是一个“微利业务”（Micro-margin business），在公司深耕的服装与鞋类类目中尤为如此，不仅利润空间受到严格挤压，还伴随着极高的退货率挑战。 — Eric Hankinson ## Mentioned Outdoor Equipped, Eric Hankinson, 亚马逊 (Amazon), Thorne, Precise Financial, Data Central, MAP Watchers, Order Desk, Ron, Hermes, Percy, 模型上下文协议 (MCP), ASICS, Thorogood, The North Face, FBA ## Video description Eric Hankinson - VP of Technology, Outdoor Equipped What if your business could make every decision the moment the data arrived, instead of waiting on a report? In high-volume operations like Amazon retail, the gap between data and decision is where margin quietly disappears — and traditional BI was never built to close it. I will walk through how I deployed AI agents and MCP servers to merge fragmented data sources across one of the largest third-party Amazon operations in the U.S., delivering per-transaction P&L, demand forecasting, and reorder decisions in real time. Attendees will leave with a practical understanding of how to architect agentic workflows for production, the pitfalls I hit along the way, and the design choices that separate demo-ware from systems that actually run a business. If you're responsible for decisions inside a data-heavy operation, this talk will show you what's already possible — and what's coming next. [Source](https://www.youtube.com/watch?v=z2IopjXwm8Y)

Source quotes

[0:04] Eric Hankinson: 一个商品页面仅对应一个通用产品代码（UPC），数据库只需记录一行数据及流入的信息即可。然而，鞋类和服装品类极其繁琐，同一款商品可能衍生出繁多的颜色和尺码组合。 [1:46] Eric Hankinson: 亚马逊的交易数据如同一个“动态生命体”，在订单提交、信用卡扣款到最终结算的不同时间节点会不断补充消费者信息。 [1:46] Eric Hankinson: 亚马逊平台有时如同“狂野西部（Wild West）”，许多夫妻小店（Mom-and-pop shops）或未授权卖家会隐匿身份倾销库存。 [4:17] Eric Hankinson: 很多人认为只要把数据直接丢给大模型并提示“随便处理这些数据并给出好建议”即可，但这种做法极其消耗 Token，且输出结果往往不够精确。 [4:17] Eric Hankinson: 团队选择将 Token 消耗前置到工具与连接器的开发阶段，大量数据清洗与聚合计算全部由前端确定性脚本和公式完成；智能体 Ron 仅作为监管与干预者存在，只在数据断流或需要介入修正时才触发模型消耗 Token。 [8:27] Eric Hankinson: 当多次出现同类查询（如按月度计算品牌退货率）时，系统会将其固化为一条“已验证路径（Verified Path）”。后续再遇到相似提问时，智能体无需再从零拆解推理，而是直接调用成熟路径，大幅降低了推理延迟与 Token 成本。 [13:12] Eric Hankinson: 通过在单个用户层面沉淀记忆，Percy 能够清晰识别提问者的岗位角色及其长期的业务目标与诉求偏好，从而针对不同员工“量身定制”符合其业务视角的回复。 [13:12] Eric Hankinson: 员工在提问时只需抛出宏观问题……系统便能综合考量复杂的市场环境、品牌特征、已知阻碍以及最新的“实时业务真相”（Live Business Truth），在底层数据随时刷新同步的前提下，迅速完成深度归因并给出精准回答。 [14:47] Eric Hankinson: 自 2012 年以来，Outdoor Equipped 始终保持着每年 30% 到 40% 的复合增长率，但借助智能体对高频分析与自动化流程的接管，企业成功将现有员工重新分配到更具战略价值的业务板块，实现了在业务体量连年高速扩张的同时无需按比例扩招团队。 [14:47] Eric Hankinson: 亚马逊平台的零售生态本质上是一个“微利业务”（Micro-margin business），在公司深耕的服装与鞋类类目中尤为如此，不仅利润空间受到严格挤压，还伴随着极高的退货率挑战。

Source quotes

Eric Hankinson: In high-volume operations like Amazon retail, the gap between data and decision is where margin quietly disappears — and traditional BI was never built to close it. Eric Hankinson: What if your business could make every decision the moment the data arrived, instead of waiting on a report?

Original transcript

[0:04] So, uh I'll start by talking just a little bit about what outdoor equipped is. Um a lot of our [0:10] uh operational challenges, uh the data foundation that we've built, [0:15] um some of the tools and um sheets that run our business along with uh the agents uh that we've [0:23] created to kind of help manage that business. So, Outdoor Equipped was founded in 2012. Um, it's [0:30] just a startup and now we're one of the largest three-piece sellers on the Amazon platform. [0:36] Our specialty category is footwear and apparel. So, a lot of these brands here um that you see, [0:42] we're basically fully exclusive, represent the brand on the platform in uh all aspects. We're [0:49] family-owned and operated. Uh so it's me uh my dad uh and three brothers uh all in various aspects of [0:56] the business. I kind of took on the technology role. Um basically what's made us grow so fast [1:03] and partner so close with these companies is our technology um focus our data foundation. Uh the [1:11] internal systems that we've built to be able to kind of share a lot of ecom knowledge uh with [1:16] these companies so they can use it internally. So you can see here an Amazon listing uh Thorn [1:24] Vitamins uh a product that I use. Um but what you can see from the listing is that it's one uh UPC. [1:32] There's only one option. Um so in terms of a database having information on this listing, [1:39] it's very simple. It's just one single row and all of the uh data that's uh flows in from it. [1:46] However, in the four and apparel category, there's a lot of challenges, right? Uh you can see here [1:53] um all the various color options and sizes. So, this particular listing actually breaks down to [1:59] 144 options. So, in terms of building out a database that allows for uh organizing data, [2:07] whether it be forecasting or marketplace data uh within this category, it's just so [2:12] much more extensive. And with that becomes our operational challenge, right? We have billions [2:18] of catalog records uh millions of new marketplace and inventory observations each day. We represent [2:24] uh 100 plus brands on the marketplace all with their own pricing, stock, margin reality to track, [2:33] 142 million monthly product records tracked, 650k uh monthly financial event transactions, [2:40] all postsale reporting. So it made it so we had to basically build a data foundation that was [2:48] very solid. Um we have a system that we named precise financial which reveals all the actual [2:55] uh financial data post sale. So we basically take all of the information from Amazon's API on a per [3:02] transaction basis live and then we apply our own internal transactions based on fulfillment method, [3:09] product size, dimensions, whatever it may be. Data central um which is basically like a live [3:17] look at the listing. So we compi we combine all of the information on the UPC's along with the [3:24] reviews that are shown on a listing. um the uh the number of sellers that are selling on the listing, [3:30] the price that they're selling at, uh impressions, whether it be from advertising data and so forth. [3:36] And then map watchers. So, in the United States, map protected product is very big, [3:40] especially with when you're dealing with very high-end premium brands like the ones I showed [3:45] you. Um they enforce MAP and Amazon can be like the wild west. Anyone can create an account. [3:52] It can be a mom and pop shop that um has some leftover inventory that they just want to sell [3:56] on the platform and hide their identity. So, Map Watchatchers, a scraping platform that allows us [4:02] to understand the competitive price factors um that are on that uh listing and how it's [4:09] going to impact their overall sales or how we can work with the brand to clean up those listings. [4:17] So, here's an example of a demand data report. So this is just for one brand. You can see the amount [4:23] of data that's here. It's information from the listing whether it be product reviews uh [4:31] impressions overall just uh sales of the listing uh our inventory and then taking an ATS from the [4:39] brand all within uh calculating a projected margin based off of shipments and movements. So we have a [4:46] bunch of segment managers that have a catalog of brands in house that they manage and this helps [4:52] them with forecasting and also understanding the overall health of a brand. Now compiling these [5:00] reports is quite extensive. It has been for uh a long time. It's six plus different sources that [5:06] we're pulling this data from. So that's where we got into uh using agents to kind of help with this [5:13] process. We developed a platform called order desk and a uh AI agent uh that monitors everything 24/7 [5:20] named Ron. Um so like I said the problem is just consolidating all those reports making sure that [5:27] everything is up to date and coming from all the various systems and making sure that they [5:32] line up into a single source of truth. Order desk allows for all those reports to be done [5:39] um ahead of time so that there is no manual work done on the front end and then the segment team [5:46] can go in and pull those reports um just directly from the system. Ron is the agent. He's actually [5:54] a Hermes agent that we built uh designed to monitor all of the uh data consolidation [6:00] and process flows along the way. making sure that if there is a breakdown, he can go in, [6:05] he can make those changes. Taking information from Amazon's API, uh, a lot of various databases from [6:12] different brands, there's always breakdowns. Uh, and Ron is there to kind of just analyze and fix [6:19] anything that's needed. So, the pipeline, so we take uh, we ingest and we consolidate the data, [6:26] pull uh, from every source and then merge into a unified uh, order per skew. validate the data, [6:33] making sure that it's all up to date, all accurate so that it is part of that source of truth, [6:39] the data report. So, as you saw with the demand data that I showed you, we have various different [6:45] color options um and just key points to kind of help the end user know exactly what they're seeing [6:53] and apply the correct filters so it's a lot easier for them to do their forecasting and then provide [6:59] a reorder suggestion. So, we wanted to make it so any segment manager that receives a file already [7:07] has a designed reorder per brand um at the given moment. So then they can just kind of understand [7:15] why is that bot suggesting that order based off of previous sales uh or inventory flow. So some [7:24] key design decisions. So a lot of people think that with when it comes to this type of data, [7:31] you just kind of create an agent uh give them the data and say, "Hey, do do with this what you will [7:37] come back with, uh what you think would be a strong recommendation." However, uh we didn't [7:42] find that very effective. Um we saw that that just ate tokens and it wasn't as accurate as we [7:49] uh would have hoped. So instead, we uh created all of the AI work on the front end to develop [7:56] the scripts that combine all of the files and then Ron uh just oversees everything. Um a few things [8:04] to make it so our token cost was more front-loaded is just making sure that the uh it's concentrated [8:12] on uh building the tool and connections rather than actually analyzing the data. It's all kind of [8:18] front-end equations. And then Ron only activates and uses tokens when it needs to intervene. [8:27] So this is a shot of what we call precise financial. This is the post sale platform [8:34] that I mentioned earlier. So this is taking each transaction that is processed through our Amazon [8:39] account live and as soon as Amazon submits that transaction it's kind of like a living organism. [8:48] They submit a portion of the data then a couple seconds later a couple minutes later they charge [8:54] the credit card and they get more consumer data and they update that information. So we update [8:59] that feed within our system along with connecting it to um our individual uh costs that aren't part [9:07] of the Amazon platform or COGS or depending on fulfillment method fulfill fulfillment [9:13] method of the various fees per brand or warehouse processing fees labeling repackaging whatever it [9:20] may be um some advertising data. However, for our segment team and our managers to really drill down [9:29] um at this data point to really do some analysis on a either per skew UPC level or [9:36] just a brand overview, um they had to go through our query builder and create a bunch of rules and [9:43] um and groups to consolidate that information and then export it into Excel to then apply the [9:51] correct calculations that they wanted too. This was efficient in the time but we saw a very good [9:58] opportunity here and that's where we uh decided to create a system called Percy. So Percy is an [10:05] integrated chatbot connected to scatter uh scattered business data through MCP servers [10:11] um with a multi- aent workflow. So I'll kind of uh discuss the breakdown of Percy uh the [10:17] message breakdown, the multi-agent layers, the quality score and feedback loop and the memory [10:22] layers that make it so successful for our internal operations. So first is breaking [10:28] down the message. So someone submits a message within the chatbot and it goes to our coordinator [10:34] agent. Um I'll use an example return rate, right? So a uh member of our team wants to [10:41] know the return rate of A6 shoes uh for the given month and how it compares for the various months [10:48] uh over the last year. The agent knows that it needs to break down this message into finding [10:54] the overall sales per month and then the overall returns apply the correct calculation um and uh do [11:03] a comparison based off of the previous months. So it's breaking down those rules into intent tasks [11:10] and evidence requirements before it submits it to these sub aents to run. So once the coordinator [11:17] agent defines the sub aent uh requirements, it pushes it to the sub aents and we have a uh [11:25] various sub aents uh depending on the data source that it's pulling from. So precise financial, [11:30] data central or map watcher. This allows us to provide a broad question um based off of the [11:40] requirements for the user and then the coordinator agent is able to then make the decision of what [11:46] data source to pull from. And then we have it go through a verification or auditor layer to be able [11:53] to verify the source that it's coming from and making sure it's up to date. But one of the most [12:00] important things is just understanding business logic with the return, right? We didn't want our [12:06] uh system to just spit back the the numbers. We wanted to provide insight and suggestions based [12:14] off of the feedback that it was providing. So we created a critique agent. So it goes to uh the [12:21] generator loop where it sends it to the critique agent. The critique agent reviews it with business [12:27] logic and scores the answer. If it comes back with a low score, it goes through the loop and tries [12:34] again. If it has a high score, it can then ship the response with confidence to the um end user. [12:44] And then there's the memory layers. So one of the most important ones is procedural memory. So use [12:51] the example of the return rate again. If we have many users that are submitting that same request [12:58] for return rate per month uh per brand over a given period of time, it defines it as a verified [13:05] path and then it makes it so the agents don't have to every time from start to finish think about [13:12] the path and the sources that it has to get that data. It remembers um that verified method and can [13:19] streamline it uh to make it a lot faster. personal memory. We have various different users, whether [13:25] it be our logistics team, our segment managers, our listings department, advertising department, [13:31] all submitting questions. So, being able to on a per user level know the role and the um overall [13:41] goal that the person consistently requests from Percy. Um, having that memory allows for them [13:48] to cater the responses to that individual user. company knowledge. So our business is considered [13:56] quite unique. Each individual brand, it's various relationships. There's different challenges within [14:02] the marketplace. Uh whether it be listing issues, fighting with Amazon themselves, pricing issues, [14:10] market condition changes, seasonal velocity, whatever it may be. And being able to kind of [14:15] frontload all of that company knowledge um allows us with all of these memory layers uh and the [14:22] overall process to have our groups basically say hey can you tell me why the sales are up or down [14:30] uh for this given brand at this moment or compared to a previous month. it can then think [14:35] uh of the various factors and uh the uh company the market conditions and company knowledge that's [14:42] built into that and then obviously just the live business truth just making sure that all systems [14:47] are up to date and always refreshed. So just like any uh AI tool um there's all these benefits [14:57] to the organization. So far, we've been able to redeploy a lot of our staff into departments that [15:04] um see that need. So, we've also been growing at a rate uh since 2012 at basically 30 to 40% per [15:12] year. Um so, being able to not have to increase headcount based off of that growth. Uh decisions [15:19] in seconds. So, Amazon is a micro margin business, especially in the footwear and apparel space. It's [15:26] a high return business. So being able to analyze everything like we do faster and faster allows us [15:35] to adjust pricing or enforce policies on the brand to be able to increase our overall margins. Um [15:42] there's also tracking within inventory based off of seasonality that requires just quick movements [15:48] on your feet uh to being able to turn things live whether it be in Amazon's warehouses or our own. [15:55] and then closer brand partnerships because of just the data that we're able to provide whether it [16:02] be AS6 thorough good the Northace uh those key partners they're able to then ingest that data [16:08] within their own systems and make design choices marketing strategies uh in inventory forecasting [16:16] based off of that information. Uh Amazon is the largest marketplace in the world right now. Um, [16:23] so being able to kind of have that big piece of the pie in their overall ecom strategy is [16:30] probably one of the strongest things that they can have right now. And that's it.