ZX.IO — AI Systems, Omnichannel Intelligence & Research Automation

01 // REPRESENTATIVE PROJECTS & ENTERPRISE TRACK RECORD

代表项目与大客户实践经验 Representative Systems & Enterprise Client Engagements

围绕“社媒大数据选号与行业 BI → 内容关联与双内核推荐产品 → 大模型业务评测与反馈架构 → AI Agent 自动化研究与研报系统”持续演进，将自然语言处理与大模型能力转化为跨国企业、知名消费品牌及营销机构可操作的生产工具与研究服务。 Evolving along a coherent technical trajectory—from social media BI and influencer valuation to dual-engine content recommendation, LLM evaluation architectures, and automated research agents—serving multinational healthcare leaders and consumer brands.

RESEARCH AGENT SYSTEM // UCRS 自动化研究与研报系统

UCRS：企业级自动化消费者研究与研报生成系统 UCRS: Automated Consumer Research & Reporting Platform

持续推进系统研发与 5.0 架构演进，整合多源数据双重筛选、大模型闭集语义分类、确定性统计分析与结构化报告生成。设计任务编排、按站点多模型路由、运行元数据留存及证据链核查机制，配套 ResearchOS 客户审阅界面，将课题 Brief 直接转化为可追溯源文件指针与统计摘要的深度研究报告。 Spearheaded ongoing R&D and v5.0 evolution of UCRS—integrating dual-stage data filtering, closed-set LLM semantic classification, statistical aggregation, and automated report synthesis with multi-model routing, execution metadata lineage, and the ResearchOS review interface.

核心角色：系统研发演进、课题版本集成、过滤与质量核查机制设计 Role: System R&D leadership, pipeline integration, filtering & quality verification

RECOMMENDATION & NLP // 内容推荐引擎与语义网

同文语义网 × 博主群体人群网：双内核内容推荐系统 Dual-Kernel Content Recommendation & Semantic Graph Engine

负责社媒内容分析与推荐引擎研发，将推荐系统与三平台语义网合并。结合 Word2Vec / Doc2Vec 语义向量、Annoy 近邻索引、博主群体关联度与专业度/收藏率（收藏÷点赞中位数）排序，支持品牌按“品牌/场景交集与并集”开展内容选题、关键词拓展与跨品类共同人群挖掘（涵盖美妆、鞋服、营养品等多品类分析实践）。 Led R&D of the social content recommendation engine, merging semantic networks with creator-group audience graphs. Combines Word2Vec/Doc2Vec embeddings, Annoy neighbor search, professionalism filtering, and save-rate ranking for cross-category topic and audience discovery.

核心角色：负责红书采集、BI 与推荐引擎研发，持续发布双内核与选题功能升级 Role: Lead R&D for RED data, BI & recommendation engine; continuous feature releases

MNC CLIENT PRACTICE // 跨国医药健康企业（晖致 Viatris）

晖致：多主题消费者洞察、需求—方案匹配与疼痛研究 Viatris: Multi-Topic Consumer Insight & Demand-Solution Research

参与肝功能、疼痛、心血管、解酒等连续多期主题研究系统（Satori 洞察生成引擎 + Dify 筛选应用 + UCRS 深度研究）研发与交付。具体改进目标人群表达、单一症状分类标准、闭集映射约束及客户指定平台运行流程；采用“生成与评价分离”及跨字段一致性校验，围绕痛点、部位×场景交叉需求与产品成分输出研究报告及共创提案。 Contributed to multi-wave research system R&D and monthly deliveries for global healthcare leader Viatris across liver health, pain management, and cardiovascular topics. Refined closed-set symptom classification standards, target-audience prompt constraints, and generation-evaluation separation.

核心角色：Satori 分类标准与提示约束改进、UCRS 疼痛需求研究集成与共创方案组织 Role: Satori classification & prompt constraint engineering, UCRS report integration & co-creation

CONSUMER HEALTH // 头部营养与母婴品牌（合生元 Biostime）

合生元：基于 UCRS 的消费者研究与分析模型迭代 Biostime: Consumer Research & Analytical Iteration on UCRS

面向知名营养健康品牌消费者研究需求，基于 UCRS 推进社媒数据筛选规则制定、系统更新与分析方法优化。针对业务取数原则与细分人群口径差异完成多轮报告修订与系统迭代，并指导业务团队运用 UCRS 完成常态化研究分析交付。 Advanced data filtering rules, UCRS system updates, and analytical model iterations for leading nutritional brand Biostime—aligning data extraction criteria with business requirements and guiding research teams on system workflows.

核心角色：数据筛选口径校准、UCRS 系统更新、分析优化版交付与业务指导 Role: Data filtering calibration, UCRS updates, analytical optimization & team enablement

MARKETING BI // 大数据选号系统与行业 BI

跨平台达人分层估值、互动质量与调性选号系统 Tiered Influencer Valuation, Comment Quality & Brand-Fit BI

为数十家品牌、跨国公司及广告服务机构研发和部署社媒分析与选号系统。以“内容分类 × 粉丝等级”二维网格计算 CPE/CPF 分位估值与套利价差，结合排除自评后的评论词频信息熵（Shannon Entropy）、评论垂直度/相关度及近远期增长异常规则，并推进内容系统与小红书、B站全站选号工作空间整合。 Built social analytics and creator selection platforms deployed across dozens of brands and agencies. Features Category×Tier CPE/CPF percentile valuation grids, Shannon Entropy comment quality scoring, semantic verticality, and cross-platform workspace integration.

核心角色：负责选号与行业 BI 核心建模及功能演进，推进内容系统与全站选号合并 Role: Core modeling for valuation & comment quality; unified content and creator selection workspaces

AICRM & AIGC // 生成式舆情、评论运营与多模态内容管线

AICRM 评论意图归因、四层状态隔离与多层内容审核系统 AICRM Comment Intent Attribution, State Isolation & Content Audit

承担 AICRM、AI 原生品牌与 AIGC 管线的架构与技术指导。将社媒评论细分为购买咨询、产品咨询、购买意向、产品评价等业务维度；设计“平台原始事实 / AI 结果 / 人工操作 / 应用状态”分层架构与客户经验包机制；在图文生成（SMCC）与审核中落地四层规则校验、同质化检测及实物参考图一致性控制。 Directed AI architecture for AICRM and multimodal content pipelines. Separates pre-sale inquiries and buyer intent from casual chatter; enforces 4-layer data state isolation (preserving human edits over AI outputs) and multi-tier compliance auditing.

核心角色：AI 开发、系统架构与技术指导，定义业务评测用例、状态分层与参考图控制 Role: AI architecture & technical direction, evaluation design, state layering & reference control

02 // SOLVING OMNICHANNEL MARKETING & RESEARCH BOTTLENECKS

为什么通用型 AI 工具难以直接解决国内多平台投放评估与下一波策略生成？ Why Generic AI Tools Struggle with China Omnichannel Spend Audit & Planning

无论是时尚鞋服、餐饮快消还是大健康品牌，在探索用 AI 审视数字营销投入（Digital Investment）并自动化产出下一波方案时，都会面临三个核心工程问题： Across fashion, food & beverage, and consumer health, brands exploring AI to audit digital spend and automate next-wave planning in China face three technical bottlenecks:

BOTTLENECK 01 // ECOSYSTEM SILOS & BASELINES

平台生态彼此封闭，仅将内部导出表格直接喂给通用大模型缺乏行业价格与语境参照。 Closed platforms block unified APIs; feeding raw spreadsheets into generic LLMs lacks category baselines.

国内小红书、抖音、微信等平台数据各自独立，外部工具无法直接拉取跨平台私密接口。而如果仅把品牌导出的投放报表上传给通用大模型，由于缺乏“内容分类 × 粉丝等级”的历史 CPE/CPF 分位基准，且没有结合笔记正文与评论区上下文，模型只能做简单的加减平均，无法判断报价是否虚高或受众是否跑偏。 RED, Douyin, and WeChat operate as isolated ecosystems. Uploading exported campaign spreadsheets directly into a generic LLM without historical Category×Tier CPE/CPF percentiles or post-and-comment context yields superficial arithmetic rather than true valuation insight.

✕ 传统局限：等待平台打通接口，或将孤立表格直接丢给无行业基准的通用对话模型 ✕ Conventional Limit: Waiting for cross-platform APIs or prompting raw sheets without category priors ✓ ZX.IO 方案：一方投放台账接入 + 社媒正文与评论联读 + 分类×粉丝等级二维估值网格校准 ✓ ZX.IO Approach: 1P sheet ingestion + post/comment joint reading + Category×Tier valuation grid

BOTTLENECK 02 // QUANTITY VS. QUALITY

同样的互动量，究竟是泛娱乐围观、模板化互动，还是真实的售前咨询与购买意向？ Does high engagement reflect entertainment onlookers and bots, or genuine pre-sale inquiries?

单看总点赞、总评论或表面 CPE 极易被偶发爆款或模板化水军误导。我们在选号与 AICRM 实践中，将数量、质量与变化拆开：用去极值指标与近远期增长对比观察稳定性，用排除自评后的评论词频信息熵（Shannon Entropy）提示模板化互动，并进一步用 NLP 区分购买咨询、产品咨询、购买意向、产品评价与对博主的泛娱乐评价。 Raw likes and surface CPE are easily skewed by outliers or repetitive bot comments. Our system separates quantity, quality, and momentum—using Shannon Entropy to flag templated comments and fine-grained NLP to distinguish pre-sale inquiries and buyer intent from creator chatter.

✕ 传统局限：把“对博主颜值的夸赞”或“抽奖口令”与“询问产品尺码/功效”混为同一互动指标 ✕ Conventional Limit: Mixing creator compliments or giveaway spam with genuine product inquiries ✓ ZX.IO 方案：评论熵丰富度分档 + 评论垂直度/相关度计算 + 售前咨询与购买意向独立归因 ✓ ZX.IO Approach: Comment entropy scoring + semantic verticality + buyer-intent attribution

BOTTLENECK 03 // FROM INSIGHT TO NEXT WAVE

如何从“复盘上一波”走向自动生成“下一波选什么人、拓什么圈、写什么 Brief”？ How to bridge post-campaign review into automated next-wave topic, audience, and brief generation?

传统复盘报告指出问题后，下一波执行仍靠人工拍脑袋。我们通过双内核推荐引擎（同文共现语义网 × 相似博主群体人群网）寻找与品牌高转化场景交集的目标人群与跨品类话题，再通过“读数据 → 分析 → 洞察 → 选题定位 → 文案框架 → 多层审核”管线，直接生成下一波可执行方案。 Instead of stopping at static post-mortem slides, our dual-kernel engine (co-occurrence semantic graph × creator-group audience graph) identifies high-affinity topics and cross-category segments, feeding directly into structured brief generation and multi-tier audit pipelines.

✕ 传统局限：只按单一关键词热度排榜，缺乏人群重合度与收藏率转化支撑 ✕ Conventional Limit: Ranking topics solely by raw keyword volume without audience overlap or save-rate proof ✓ ZX.IO 方案：品牌/场景并集与交集向量检索 + 收藏率排序选题 + 结构化创作与审核闭环 ✓ ZX.IO Approach: Union/Intersection vector retrieval + save-rate topic ranking + structured brief pipeline

STAGE 01 // DATA & STATE LAYER

一方台账与多平台社媒联读 1P Sheets & Social Post-Comment Joint Ingestion

结合品牌一方投放报表与小红书/抖音/B站笔记正文及评论流；区分平台原始事实、AI 结果与人工修正状态。 Combines brand 1P campaign sheets with social posts and comments while isolating raw platform facts, AI outputs, and human edits.

STAGE 02 // GRID & ENTROPY AUDIT

网格估值校准与评论意图拆解 Grid Valuation & Comment Intent Decomposition

按分类×粉丝等级计算 CPE/CPF 分位与估值价差；利用评论熵、垂直度及细粒度 NLP 分离购买咨询与泛娱乐互动。 Computes Category×Tier CPE/CPF percentiles and applies comment entropy and NLP to isolate pre-sale inquiries and buyer intent.

STAGE 03 // DUAL-KERNEL RECO

语义网×人群网下一波策略推演 Dual-Kernel Next-Wave Topic & Creator Matching

通过关键词/笔记/博主关系矩阵与向量距离，挖掘跨品类共同兴趣人群，按专业度与收藏率筛选下一波候选矩阵。 Uses keyword-post-creator relation matrices and vector correlation to uncover cross-category audiences and high save-rate topics.

STAGE 04 // UCRS & BRIEF DELIVERY

可追溯研究报告与创作审核闭环 Traceable UCRS Reports & Audited Briefs

经 UCRS 自动化研究管线输出带证据链的策略报告，并结合客户经验包生成文案框架与四层规则审核报告。 Synthesizes evidence-linked UCRS research reports and generates structured briefs validated against 4-tier compliance rules.

03 // INTERACTIVE MULTI-SECTOR DECISION SANDBOX

交互式多行业决策沙盒：投放估值校准、评论意图拆解与下一波推荐 Interactive Multi-Sector Sandbox: Valuation Audit, Intent Analysis & Next-Wave Recommendation

本沙盒基于我们研发的大数据选号估值网格、评论信息熵/垂直度分析与双内核推荐引擎方法构建。您可以点选切换不同行业场景（时尚鞋服与配饰 / 餐饮快消与咖啡 / 美妆个护与营养健康），并点选任意投放波段卡片查看算法诊断逻辑： Built upon our Category×Tier Valuation Grid, Comment Entropy/Verticality Analysis, and Dual-Kernel Recommendation Engine. Tap any industry scenario below and select a campaign flight card to inspect the diagnostic workflow:

点选切换行业演示场景： SELECT INDUSTRY SCENARIO:

方法演示样本：时尚鞋服与配饰 · 季度新品多平台投放审计（点选下方任意波段查看网格分位与评论意图诊断） Methodology Sample: Fashion & Accessories · Multi-Platform Spend Audit (Tap any flight below to inspect)

METHODOLOGY DEMO // SYNTHETIC BENCHMARK

投放波段 / 达人分层矩阵 Campaign Flight / Creator Tier	预算占比 Spend Share	网格 CPE 分位 Grid CPE Pct	评论熵(丰富度) Comment Entropy	售前咨询/意向率 Pre-Sale Intent	策略建议 Recommendation
A. 小红书 · 职场通勤与穿搭垂类腰部博主 A. RED · Office Commuter & Styling Mid-Tier Creators	28%	¥14.2 (P25 优质)	4.32 bits (高)	24.6%	SCALE +35%
B. 短视频平台 · 泛娱乐颜值变装头部达人 B. Short-Video · Entertainment Transition Top Creators	36%	¥9.8 (有效 P82)	1.85 bits (低)	3.1%	CUT -50%
C. 小红书 · 时尚街拍与多单品合集头部博主 C. RED · Street-Style & Multi-Item Compilation KOLs	22%	¥18.5 (P50 中位)	2.64 bits (中)	11.4%	RE-BRIEF
D. 运动休闲与都市生活方式跨界KOC矩阵 D. Cross-Category Athleisure & Lifestyle KOC Matrix	14%	¥12.6 (P20 洼地)	4.45 bits (高)	28.2%	SCALE +30%

A. 小红书 · 职场通勤与穿搭垂类腰部博主（分类×粉丝等级网格评估） A. RED · Office Commuter & Styling Mid-Tier Creators (Grid Valuation)

分类×量级 CPE 处于 25% 分位（性价比优于同网格 75% 样本） · 收藏/点赞中位数：68% Category×Tier CPE at 25th Percentile (Top quartile value) · Median Save/Like Ratio: 68%

评论垂直度（评论与品类核心词组语义相关超阈值比例） Comment Verticality (Category Semantic Relevance Ratio) 82%

售前咨询与购买意向占比（AICRM 意图分类：问尺寸/自重/货号） Pre-Sale Inquiry & Purchase Intent Share (AICRM NLP) 24.6%

收藏率表现（收藏数 ÷ 点赞数中位数，衡量长尾种草心智） Save-to-Like Median Ratio (Long-Tail Reference Value) 68%

算法诊断：按“内容分类×粉丝等级”网格校准，该波段报价低于模型中位估值（存在正价差）。排除博主自评后评论信息熵达 4.32 bits，且 AICRM 意图模型识别出高频产品咨询（如“电脑能不能装下”、“不同身高背带长度”、“耐磨度”），而非泛娱乐围观。建议下一波扩充预算并复用该痛点词簇。 Algorithmic Diagnosis: Within the Category×Tier grid, actual spend sits below median valuation (positive arbitrage spread). Comment entropy reaches 4.32 bits after excluding self-replies, with AICRM intent models flagging dense pre-sale product inquiries rather than generic chatter. Recommend scaling budget.

基于“同文共现语义引擎 × 博主群体人群引擎”生成的下一波选题与选人方案 Next-Wave Topic & Creator Plan Powered by Dual-Kernel Recommendation Engine

DUAL-KERNEL RECO // WORD2VEC + ANNOY + AUDIENCE OVERLAP

STEP 01 // GRID VALUATION & FILTERING

基于网格估值与互动质量的预算重配 Grid Valuation & Quality Reallocation

缩减泛娱乐低垂直度波段（-50%）：表面互动量虽大，但评论垂直度与产品咨询占比极低，大多数互动属于对博主的评价而非本品。 Reduce Low-Verticality Entertainment Flights (-50%): High raw views but minimal product relevance; comments target the creator rather than the brand.
增投高收藏率腰部垂类矩阵（+35%）：筛选分类×粉丝网格中 CPE 低于 50% 分位、且近期互动增长无跳变异常的创作者。 Scale High Save-Rate Mid-Tier Verticals (+35%): Filter creators with CPE below the 50th grid percentile and stable growth trajectories.
利用 CPF 网格补估挖掘潜力新号：对尚无公开刊例价但垂直度位于前四分位的上升期博主进行估值比价。 Discover Rising Creators via CPF Grid Imputation: Value emerging high-verticality creators lacking rate-card history using tier CPF benchmarks.

STEP 02 // AUDIENCE & TOPIC EXPANSION

双内核引擎挖掘的跨品类人群与话题交集 Dual-Kernel Cross-Category Audience & Topic Discovery

人群引擎跨品类发现 #1：都市网球 / 轻户外运动群体 — 相似博主群体关联度高（相对排名前 10%），目标受众在运动穿搭与通勤单品间存在强交集。 Audience Engine Discovery #1: Urban Tennis & Light Outdoor — High creator-group vector correlation (top decile rank) with strong buyer overlap.
语义引擎同文共现拓词 #2：“差旅登机 / 轻量化通勤 / 一包多穿” — 通过关键词与笔记向量交集（`&`）检索，该场景词簇下的收藏/点赞中位数显著高于大盘。 Semantic Co-occurrence Cluster #2: “Business Travel & Lightweight Commute” — Intersection (`&`) vector queries reveal superior save-to-like ratios.
多品牌态度独立归因对比：结合笔记与评论联读，剥离合集帖中其他配饰干扰，锁定本品正评与核心卖点共现词。 Multi-Brand Independent Attitude Binding: Jointly reads posts and comments to isolate brand-specific sentiment in multi-item compilations.

STEP 03 // BRIEF & MULTI-TIER AUDIT

数据驱动创作 Brief 与四层规则审核 Data-Driven Creative Brief & 4-Tier Content Audit

痛点与功能导向选题框架：按“读数据 → 洞察痛点 → 选题定位 → 文案框架”生成 Brief，将评论区高频售前疑问（容量、尺寸、上身比例）前置入正片。 Pain-Point Driven Brief Structure: Embeds top pre-sale comment inquiries (capacity, sizing, proportion) directly into the creative hook.
同质化检测与四层审核校验：在发布前按“平台通用、品类、品牌产品、单条 Brief”四层标准自动检查必含信息与文案同质化风险。 Homogeneity & 4-Tier Compliance Check: Audits drafts across Platform, Category, Product, and Brief rules while flagging repetitive copy structures.
评论区上下文组批运营：投后接入 AICRM，结合笔记正文与客户经验包生成高商业友好度的咨询回复与维护建议。 Context-Aware Comment Operations: Post-launch AICRM batches comments per post with brand experience packs for commercially aligned replies.

如何将品牌一方表格、多平台社媒内容与大模型分析结合为可维护的生产系统？ How We Combine Brand 1P Sheets, Multi-Platform Social Data & LLM Pipelines

AICRM STATE LAYERING // UCRS EVIDENCE TRACEABILITY

数据层 A：品牌一方投放台账与多源字段对齐 Layer A: Brand 1P Sheets & Field Normalization 1P & BI

针对不同平台导出表字段不一、缺失或口径差异（如蒲公英、主页、IP 地域字段优先级治理，报价缺失时的 CPF 网格补估），建立标准化的 ETL 与指标清洗规则。 Normalizes disparate exported sheets with explicit field-priority governance (e.g., Pugongying > Profile > IP location rules) and CPF grid rate imputation.

数据层 B：笔记正文 + 评论联读与闭集语义打标 Layer B: Post-Comment Joint Reading & Closed-Set Tagging NLP & SATORI

同一篇笔记下的评论共享笔记上下文组批处理；对多品牌/多产品提及进行独立态度绑定，并采用两阶段闭集分类体系（先建立领域标准，再批量映射至每条样本），避免通用模型随意发散。 Batches comments with shared post context, binds brand-specific attitudes independently, and applies two-stage closed-set taxonomy mapping (Satori/UCRS).

数据层 C：客户经验包与请求级 RAG 分层上下文 Layer C: Client Experience Packs & Request-Level RAG EXPERIENCE PACK

将全局模板、按请求检索的 RAG 知识与按客户/任务沉淀的“经验包”分离；生成与审核任务解耦，支持批量前更新策略、批内固定缓存。 Separates global templates, request-level RAG retrieval, and client-specific Experience Packs while decoupling generation from verification agents.

══ 架构治理 ══►

四类事实状态分层与 UCRS 证据链追溯 4-State Fact Layering & UCRS Evidence Lineage TRACEABLE

1. 平台原始事实：保留原始图文、评论、互动计数与源文件指针；
2. AI 分析结果：保存生成模型、策略模板哈希、分类标签与结构化理由；
3. 用户人工操作：人工修正与业务审核优先级高于 AI 结果，防止下次数据同步覆盖人工校准；
4. 应用展示与报告：通过 UCRS 与 ResearchOS 输出带统计底表与代表原文引用的分析报告。 1. Platform Raw Facts: Preserves source posts, comments, metrics, and file pointers;
2. AI Analytical Outputs: Logs model routing, prompt strategy hashes, and structured tags;
3. Human Operations: Human edits take strict precedence over AI outputs across sync cycles;
4. Application & ResearchOS: Delivers statistical tables and verbatim quotes for executive review.

✓ 将运营维护评论与自然消费者反馈分开统计，确保投后评估与消费者洞察不被自身水军或维护帖污染。 ✓ Explicitly separates brand-maintained comments from organic consumer feedback to keep post-campaign analytics uncontaminated.

04 // CORE TECHNICAL ARCHITECTURE & CAPABILITIES

四大核心技术底座（源自真实研发代码、产品手册与连续课题交付） Four Core Technical Pillars Grounded in Production Engineering

不堆砌空泛的大模型口号，每一项能力均有明确的数据建模逻辑、系统实现机制与业务落地验证： Built on concrete statistical modeling, NLP vector architectures, and verifiable multi-agent research workflows:

SYS-01 // UCRS & SATORI RESEARCH ENGINE AUTOMATED RESEARCH & INSIGHT

UCRS 自动化研究系统与两阶段闭集洞察管线 UCRS Automated Research & Two-Stage Closed-Set Insight Pipeline

将社媒数据采集筛选、领域分类标准生成、批量语义打标、交叉统计分析与研报撰写串联为可复核的 Agent 研究工作流。 Connects social data filtering, domain taxonomy generation, batch semantic tagging, cross-tab statistics, and report synthesis into a traceable workflow.

两阶段闭集分类（Satori / UCRS）：先从样本集合建立业务分类体系与单一症状/需求标准，再将同一标准并发映射至每条数据，杜绝逐条随意发散。 Two-Stage Closed-Set Classification: Establishes unified domain taxonomies first, then maps every record against strict closed-set constraints.
生成与评价分离 & 跨字段一致性核查：独立检验核心产品相关性、去重及“标题—人群—需求—原文”逻辑一致性。 Decoupled Generation & Evaluation: Validates product relevance, deduplication, and cross-field consistency across title, audience, need, and quote.
多模型路由与运行元数据追溯：完整记录源文件指针、提示模板哈希与批次重试日志，配套 ResearchOS 支持客户审阅。 Multi-Model Routing & Metadata Lineage: Tracks source pointers, prompt template hashes, and batch retry logs with ResearchOS review.

应用实证：已用于晖致（肝功能/疼痛/心血管/解酒）、合生元等跨国及头部品牌的连续课题研究与报告交付。 Proven Application: Deployed across multi-wave consumer and medical insight engagements for Viatris and Biostime.

SYS-02 // DUAL-KERNEL RECOMMENDATION ENGINE SEMANTIC × AUDIENCE GRAPH

内容推荐引擎与三平台语义网（语义×人群双内核） Dual-Kernel Content Recommendation & Tri-Platform Semantic Graph

改变仅靠单一关键词热度排榜的传统做法，同时测量内容之间、人群之间以及词与人之间的关联关系，将数据分析变为可操作的选题与选人工具。 Moves beyond raw keyword popularity by modeling relationships across content, creator communities, and brand scenarios.

同文共现语义网 vs. 相似博主群体人群网：前者挖掘垂类深度内容词簇，后者通过博主向量聚合与 correlation 距离寻找跨品类共同兴趣人群。 Co-Occurrence Semantic Net vs. Creator-Group Audience Net: Combines vertical keyword depth with cross-category audience affinity discovery.
并集（`,`）与交集（`&`）组合向量检索：基于 Word2Vec / Doc2Vec 与 Annoy 近邻索引，支持“品牌 × 场景 × 痛点”多维组合查询。 Union & Intersection Vector Retrieval: Word2Vec/Doc2Vec embeddings and Annoy indices supporting complex Brand × Scene × Pain-Point queries.
业务指标召回 + AI 定性筛选 + 收藏率排序：结合专业度过滤与收藏率（收藏数÷点赞数中位数）排序，输出高转化潜力选题。 Metric Recall + AI Qualitative Filter + Save-Rate Ranking: Filters candidates by creator professionalism and median save-to-like ratios.

应用实证：由张鑫主导研发并持续发布升级，已形成覆盖美妆、个护、鞋服、营养品等多品类的拓词、定位与联名选号能力。 Proven Application: Led and continuously shipped by Zhang Xin across beauty, footwear, personal care, and nutritional brand scenarios.

SYS-03 // BIG DATA CREATOR VALUATION & BI GRID PRICING & ENTROPY

大数据选号系统：二维网格估值与评论信息熵分析 Big Data Creator Selection: 2D Valuation Grid & Comment Entropy

将博主报价、历史表现与互动质量拆解为可计算的统计基准，解决不同类目、不同粉丝量级达人无法直接横向比价的难题。 Decomposes creator pricing, historical momentum, and comment quality into comparable statistical benchmarks across tiers.

分类×粉丝等级二维估值网格：图文与视频分开建模，按网格计算 CPE（报价÷近期互动中位数）的 25%/50%/75% 分位，缺失报价由网格 CPF×粉丝数补估，量化估值价差。 Category × Follower-Tier Valuation Grid: Computes 25th/50th/75th CPE percentiles separately for图文 and video, imputing missing rates via grid CPF.
排除自评的评论信息熵（Shannon Entropy）：按长度裁剪、分词计算以 2 为底的词频熵并标准化分档，联动原始评论检查模板化与重复水军互动。 Shannon Entropy Comment Scoring: Excludes author self-replies and computes base-2 lexical entropy to flag repetitive or templated engagement.
评论垂直度、相关度与异常检测：结合近远时间窗口增长对比、收藏/评论比例跳变规则及疑似虚增粉丝识别，保护价格基准可靠性。 Verticality & Anomaly Detection: Combines near/far growth windows and save/comment jump rules to safeguard valuation baselines.

应用实证：已在小红书、B站等多平台选号与行业 BI 中落地，并完成内容系统与全站选号工作空间整合。 Proven Application: Deployed across RED and Bilibili creator selection and category BI workspaces.

SYS-04 // AICRM & MULTIMODAL CONTENT PIPELINE INTENT NLP & CONTENT AUDIT

AICRM 生成式舆情运营与参考图驱动多模态管线 AICRM Generative Comment Ops & Reference-Driven Multimodal Pipeline

围绕商业友好度、上下文隔离、人工反馈保护与产品一致性，构建企业级评论运营、图文生产（SMCC）与多层内容审核体系。 Engineered around commercial alignment, context isolation, human-edit preservation, and strict product reference consistency.

细粒度评论意图与主体归因：区分购买咨询、产品咨询、购买意向、产品评价与对博主的评价；区分本品与竞品、自然反馈与维护评论。 Fine-Grained Intent & Entity Attribution: Separates pre-sale inquiries, buyer intent, and product feedback from creator compliments.
四层内容审核与同质化判别：按通用/平台、品类、品牌产品、单条 Brief 分层检查必含信息与风险，并对多篇文案做观点/结构同质化检测。 4-Tier Content Audit & Homogeneity Detection: Checks mandatory elements across Platform, Category, Product, and Brief levels.
实物参考图优先与可观察多模态流程（SMCC）：生成前核对正/侧实物参考图与最终 Prompt，分离外观实拍与空间尺寸约束，支持并发生成与失败槽位局部重试。 Reference-First Multimodal Control (SMCC): Enforces verified front/side product reference images and slot-level retry state machines.

应用实证：已形成 AICRM 多平台评论运营、Coze 四维判别评测原型与 SMCC 图文流水线工程实现。 Proven Application: Implemented across AICRM comment operations, 4-dimension evaluation workflows, and the SMCC pipeline.

05 // PRINCIPAL PROFILE & CAREER TRAJECTORY

关于张鑫（Zhang Xin） About Zhang Xin

AI 技术负责人。拥有十余年软件研发与技术团队管理经验，聚焦大模型、自然语言处理与 AI Agent 系统建设。经历涵盖一线互联网大厂（腾讯、美团）、自主创业（极乐实验室）及公司被收购后担任蕃茄蛋 CTO。 AI Technology Leader with over a decade of software engineering and technical team management experience, focusing on LLMs, Natural Language Processing, and AI Agent systems across Tencent, Meituan, entrepreneurial founding, and CTO leadership.

张鑫 · Zhang Xin Zhang Xin

AI 技术负责人 · 蕃茄蛋 CTO · 极乐实验室创始人 AI Technology Leader · CTO at Fanjiedan · Founder of Jile Lab

西安电子科技大学 · 本科 · 工学学士（2008–2012） B.Eng. in Information Countermeasure Technology, Xidian University (2008–2012)

拥有十余年软件研发与技术团队管理经验，早年曾任腾讯软件工程师、美团技术经理；2018 年起自主创业创办极乐实验室，深耕社交媒体自然语言处理与数据分析系统研发，后公司被蕃茄蛋收购并担任 CTO，带领技术与产品团队为数十家品牌、跨国公司及机构研发和部署社媒分析、内容推荐与 AI 应用系统。曾受邀于长江商学院等机构分享《从 GPT2 到 AI Agent：应用层 2B 实战经验》，持续将业务数据、语义建模与大模型 Agent 工程转化为可落地、可核查的企业级产品。 Over a decade of software R&D and engineering leadership. Former Software Engineer at Tencent and Engineering Manager at Meituan. Founded Jile Lab in 2018 to build social media NLP and data analytics products; following the company’s acquisition by Fanjiedan, serves as CTO leading technical and product teams across social BI, recommendation engines, and enterprise AI agent platforms.

三大核心技术能力支柱： Three Core Technical Competencies:

1. 大模型与自然语言处理：围绕业务任务开展模型选型与适配、语义表示、信息抽取及结构化输出，结合评测和规则校验构建可复用数据处理能力；
2. AI Agent 与整体系统架构：统筹数据接入、模型路由、任务编排、工具集成与报告生成，设计证据追溯、任务恢复和质量检查机制；
3. 技术路线与团队管理：将业务目标转化为技术路线、模块分工与交付计划，具备研发团队管理、技术培训及跨业务共创协作经验。 1. LLMs & NLP: Task-specific model adaptation, semantic representation, information extraction, and structured evaluation;
2. AI Agents & System Architecture: End-to-end orchestration across data ingestion, multi-model routing, evidence lineage, and quality verification;
3. Technical Strategy & Team Leadership: Translating business goals into architecture roadmaps, engineering standards, and client delivery.

蕃茄蛋 | CTO Fanjiedan (Tomatogo) | CTO ACQUISITION – PRESENT

带领技术与产品团队研发社媒数据分析、红书 BI、双内核内容推荐引擎与 AI 应用系统；承担 AICRM、AI 原生品牌与 AIGC 管线的 AI 开发、系统架构与技术指导，持续推进 UCRS 自动化研究与研报系统升级及跨国企业课题交付。 Leading technical and product teams building social data analytics, RED BI, dual-kernel content recommendation engines, AICRM, and the UCRS automated research platform.

极乐实验室 | 创始人 Jile Lab | Founder 2018 – ACQUISITION

2018 年起自主创业，研发社交媒体自然语言处理、达人分层估值选号与数据分析系统，为品牌主、广告公司与研究机构提供数据产品服务；后公司被蕃茄蛋收购。 Founded in 2018 to build social media NLP, creator valuation, and marketing analytics systems for brands and agencies; subsequently acquired by Fanjiedan.

美团 | 技术经理 Meituan | Engineering Manager 2015 – 2018

先后参与北京大交通（火车票预订）与上海智慧餐厅 SaaS 研发；在大交通阶段管理约 10 人前端团队，在智慧餐厅阶段带领约 20 人技术团队，推进自动化测试、工程质量规范与团队培养。 Participated in Meituan Transportation (train ticketing) and Smart Restaurant SaaS R&D; led ~10-person frontend team and later ~20-person engineering team in the Smart Restaurant phase.

腾讯 | 软件工程师 Tencent | Software Engineer 2013 – 2015

参与腾讯风铃移动端建站产品及相关互动技术研发，推进前端技术规范建设与内部技术培训。 Developed Tencent Fengling mobile site-building products and interactive web technologies, advancing frontend engineering standards and internal training.

西安电子科技大学 | 本科 · 工学学士 Xidian University | B.Eng. 2008 – 2012

信息对抗技术专业，四年制本科，获工学学士学位。 Bachelor of Engineering in Information Countermeasure Technology.

06 // FDE COLLABORATION & PILOT WORKFLOW

如何开启合作？—— 面向具体业务课题的快速共创与系统落地 How We Collaborate: Problem-Driven Co-Creation & FDE Deployment

无需漫长的传统 IT 立项。我们可以围绕品牌过往 1–2 波战役的投放台账、特定品类消费者研究命题或内容生产审核痛点，快速打通数据、校准标准并交付可验证的分析与工作流产物： Start with a concrete business proposition—such as auditing 1–2 past campaign datasets, running a category demand-solution study, or deploying an audited content workflow:

PHASE 01 // CASE ALIGNMENT & DATA FILTERING

阶段一：业务用例对齐与多源数据清洗 Phase 1: Use-Case Alignment & Data Filtering

明确品牌核心业务问题（如投放性价比校准、品类痛点与场景研究），接入一方投放报表与目标平台社媒笔记/评论数据，完成品牌别名、品类过滤与分类标准对齐。 Define core business questions, ingest 1P campaign sheets alongside target social posts and comments, and align closed-set classification standards with brand teams.

产出：《课题分类标准契约、数据筛选清单与网格价格参照基线》 Output: Classification Taxonomy Contract & Category Valuation Baseline

PHASE 02 // AUDIT & INSIGHT ENGINE

阶段二：投放质量审计与消费者需求—方案洞察 Phase 2: Spend Quality Audit & Demand-Solution Insight

运行二维网格估值、评论信息熵与细粒度意图分类模型，剥离泛娱乐水分，量化真实售前咨询与产品反馈，并经 UCRS 管线输出带原文证据与统计检验的深度报告。 Execute grid valuation, comment entropy scoring, and fine-grained intent classification to separate vanity views from buyer inquiries, generating evidence-linked UCRS reports.

产出：《多平台投放质量审计 / 消费者痛点与场景交叉研究报告》 Output: Multi-Platform Spend Audit & Cross-Scenario Consumer Insight Report

PHASE 03 // NEXT-WAVE RECO & WORKFLOW

阶段三：下一波选题选人推荐与工作台赋能 Phase 3: Next-Wave Recommendation & Workflow Enablement

基于“同文语义网 × 博主群体人群网”双内核输出下一波高关联人群、跨品类拓词与达人候选表，沉淀客户专属经验包，并指导业务团队常态化使用分析与内容工作流。 Generate next-wave cross-category topics and creator candidates via the dual-kernel engine, build client-specific Experience Packs, and enable internal teams on the workflow.

产出：《下一波选题与选号推荐方案》+ 客户经验包与系统操作赋能 Output: Next-Wave Topic/Creator Playbook + Experience Pack & Team Enablement

将社交媒体数据、语义分析与大模型，转化为可追溯的商业决策与增长系统。 Turning Social Media Data, Semantic Modeling & LLMs into Auditable Marketing & Research Systems.