Yancy

项目背景Background

前身 QueryForge（Text2SQL 数据分析）。按内部架构扩展为三个产品共享的 AI 核心：AI Analyst（数据分析）/ AI Agent（自动执行）/ AI Customer（客服销售）。三者共用同一套 LLM / RAG / Text2SQL / Workflow，而不是各自重复用户系统、Agent、权限、日志、数据库和知识库。

Grew out of QueryForge (Text2SQL). One shared AI core for three products: Analyst (analytics), Agent (execution), Customer (support/sales). Shared LLM / RAG / Text2SQL / Workflow — not three copies of users, agents, RBAC, logs, DBs, and knowledge bases.

技术架构Architecture

5 层：可编辑安装的共享 Python 包 + Java 统一认证网关 + Nginx 入口 + Go Worker（Kafka / 通知 / 定时）+ 三个 FastAPI 后端 + 三个 Next.js 前端。CI：pytest + mvn test + tsc --noEmit；本地 Go / Python / Java 测试与 compose config 均通过。

Five layers: editable-installed shared Python package + Java auth gateway + Nginx + Go worker (Kafka / notify / cron) + three FastAPI backends + three Next.js UIs. CI: pytest + mvn test + tsc --noEmit; local Go / Python / Java tests and compose config green.

职责范围What I owned

• 🏗 platform-common — 把三产品重复的数仓目录 / 连接器 / LLM 客户端 / SQL 校验 / 认证 / Redis 收成共享 pip 包。数据源注册表默认按进程目录落盘，收成共享包后会误共享「谁注册了哪个源」；用独立环境变量把注册库拆开（真实数仓结构仍共享）。Collapsed duplicated catalog / connectors / LLM client / SQL validator / auth / Redis into a shared pip package. Datasource registry defaulted to process cwd, so three products accidentally shared registrations — split with an explicit env var (warehouse catalogs still shared).
• 🔐 三产品统一 Java 认证网关 + SSO；前端共用同一 JWT，禁止浏览器直连 Python 服务。Single Java auth gateway + SSO; one JWT; no browser-to-Python bypass.
• 🚦 Nginx 统一入口 + IP 限流；可切换 TLS。Nginx front door + IP rate limit; optional TLS.
• 📨 Go Worker 消费通知类 Kafka topic（SMTP + Telegram）；其余 topic 只写回执（业务已在 Python 完成）；cron 由 Go 触发、Python 执行。Go worker: real SMTP + Telegram on the notification topic; other topics write receipts only (Python already ran the work); cron = Go WHEN, Python HOW.
• 🗄 Redis 用于会话 / Agent 状态 / 租户限流 / 审批锁。未配置时诚实 no-op；审批竞态另有数据库原子更新兜底。Redis for session / agent state / tenant rate limit / approval lock. Honest no-op when unset; DB atomic update as last race guard.
• 🧩 领域模型覆盖项目 / Agent / 工作流 / 任务；套餐配额与执行类权限可调。Domain: project / agent / workflow / task; plan quotas and execution permissions are patchable.
• 🤝 统一对话入口按产品上下文路由到各 Python AI Gateway，并保留各产品原有 API 形状。Unified chat entry routes by product context to each Python AI gateway, keeping existing API shapes.
• 🎛 Prompt / 模型路由 / 租户限流与成本边界；高风险工具 HITL 审批；日志审计与灰度。日常用 Cursor / Claude Code 做重构、补测与排障，不盲信生成代码。Prompt / model routing / tenant rate limits and cost caps; HITL for high-risk tools; audit logs and canaries. Daily Cursor / Claude Code for refactor / tests / triage — generated code is reviewed, not trusted blindly.
• ✅ 诚实披露：Analyst 对接生产数仓；Agent 有 Tool+Workflow 与高风险人工审批；Customer 混合语义检索 + 退款审批幂等。本机 Docker 端到端通常未拉起，不冒充已验证。Honest status: Analyst on a live warehouse; Agent Tool+Workflow with HITL for high-risk ops; Customer hybrid search + idempotent refunds. Local Docker E2E usually not up — not claimed as proven.

项目背景Background

基于阿里云 MaxCompute 的实时风控决策平台，<300ms 响应支撑交易前干预。规模：16 风控决策模块 / 53 API / 16 可视化 Panel / 27 热点 Job 并行预热，日均 4.2 亿订单事件 + 7.37 亿广告点击，跨两个 ODPS 项目接入核心订单与行为表。

MaxCompute-backed realtime risk platform, <300ms for pre-trade intervention. Scale: 16 decision modules / 53 APIs / 16 panels / 27 warmup jobs, 420M daily order events + 737M ad clicks, core order and behavior tables across two ODPS projects.

职责范围What I owned

• 🔧 ODPS 查询层封装（跨 project + schema 混合查询），处理多 BU 联合视图 + 权限拒绝兜底 + Tunnel 重连。PyODPS query layer across projects/schemas, multi-BU union views, permission fallback, Tunnel reconnect.
• ⚡ MaxCompute SQL 性能调优 — 把多个 ADS 任务从 4h+ 优化到 <30 min（4-tier UNION ALL / COLLECT_LIST+LATERAL VIEW / COUNT(*) 替代 COUNT DISTINCT / ROW_NUMBER 去重 / MAPJOIN / skewindata / 整数比较替代 DATEADD / ten×ten cross join 替代 N 行 UNION ALL）。MaxCompute SQL: several ADS jobs from 4h+ to <30 min (4-tier UNION ALL, COLLECT_LIST+LATERAL VIEW, COUNT(*) vs DISTINCT, ROW_NUMBER, MAPJOIN, skewindata, integer compare vs DATEADD, ten×ten vs N-way UNION ALL).
• 🕸 NetworkX 知识图谱 + 风险评分引擎 — 识别可疑交易 / 作弊广告点击 / 异常存提 / KOL 网络拓扑 / collusion clustering / multi-account 关联。NetworkX knowledge graph + risk scoring: suspicious trades, ad-click fraud, deposit/withdraw anomalies, KOL topology, collusion clustering, multi-account links.
• 🚀 27 热点 Job 并行预热 — 冷启从 600s+ 压到 60–120s，30min TTL 缓存命中 <300ms。27 parallel warmup jobs: cold start 600s+ → 60–120s; 30min TTL cache hits <300ms.
• 📊 16 可视化 Panel — Risk Radar / Transaction Pattern / Anomaly Stream / KOL Network / Wool Hunter (XGBoost AUC 0.967) / AML 等。16 panels — Risk Radar / Transaction Pattern / Anomaly Stream / KOL Network / Wool Hunter (XGBoost AUC 0.967) / AML.
• ⚠ Flink 实时风控 — Java DataStream + CEP：5 min 窗口连续 10 笔 same price+amount 触发警报，RocksDB state + 增量 checkpoint + externalized retention。Flink DataStream + CEP: 10 same-price+amount trades in 5 min fires an alert; RocksDB + incremental checkpoint + externalized retention.
• 🛠 DataWorks 调度治理：编写 _di / _wi / _mi 三粒度留存表 INSERT，以及 SelectDB/Doris DDL 转换。DataWorks: _di/_wi/_mi retention INSERT scripts plus SelectDB/Doris DDL conversion.

项目背景Background

面向 DashFlix 视频产品的 AI 内容理解、推荐与多模态审核平台。820 部剧集 / 145,967 行 catalog，统一视频帧+字幕+弹幕+用户行为到向量空间。规模：17 后端模块 / 16 前端组件 / 52 API / 18 前端 Tab / 26 dashboard 并行预热 98.3s 完成，首屏后 <50ms 缓存命中。

AI understanding, recommendation, and multimodal moderation for DashFlix. 820 titles / 145,967 catalog rows; frames + subtitles + comments + behavior in one vector space. Scale: 17 backend modules / 16 UI components / 52 APIs / 18 tabs / 26 dashboards warmed in 98.3s, then <50ms cache hits.

职责范围What I owned

• 🎬 多模态内容理解管道 — 视频帧抽样 + PaddleOCR + Whisper ASR + LLM 剧情摘要，统一进 pgvector 768 维向量空间。Multimodal pipeline — frame sampling + PaddleOCR + Whisper ASR + LLM summaries into pgvector 768-d.
• 🔍 RAG 检索系统 — pgvector 语义召回 + Cohere Rerank，「找类似剧」Top-10 召回准确率 80%+，P95 延迟 <800ms。RAG — pgvector recall + Cohere Rerank; similar-title Top-10 accuracy 80%+, P95 <800ms.
• 🤖 LangGraph 多代理工作流 — lang_detect / conversion / recommend / dialog 4 代理，剧情问答 8 秒出答案。LangGraph 4-agent workflow (lang_detect / conversion / recommend / dialog); plot Q&A in 8s.
• 📈 用户行为分析模块 — ODPS 日级用户事件 → dashboard 缓存层，26 dashboard 并行预热。Daily ODPS events → dashboard cache; 26 parallel warmups.
• 🐳 Docker Compose — PostgreSQL pgvector + Redis + FastAPI + Next.js.

项目背景Background

集团评估接入加密货币充提通道（跨链 USDT/USDC）面临的欺诈/洗钱/钓鱼地址风险。需要 Go 高并发网关 + AI 风控 Agent + 图谱风控的实时拦截系统，且在 Neo4j/Milvus/Kafka 不可用时优雅降级（fail-closed 设计）。

Evaluating crypto deposit/withdraw rails (cross-chain USDT/USDC) against fraud, AML, and phishing. Needed a Go high-concurrency gateway + AI risk agents + graph risk, fail-closed when Neo4j/Milvus/Kafka are down.

职责范围What I owned

• 🔗 Go 高并发风控网关 — gRPC + OpenTelemetry 全链路追踪，幂等 Pipeline 统一 8 个调用点，Bloom Filter 三层校验（内存 + Redis + 定期重建），Kafka 事件总线解耦。Go gateway — gRPC + OpenTelemetry, idempotent pipeline at 8 call sites, 3-layer Bloom (memory + Redis + rebuild), Kafka bus.
• 🕸 AML / Behavior / Phishing 三 Agent 并行风控 — Neo4j GraphRAG 识别 peeling / mixer / layering 洗钱模式，Milvus 向量检索比对钓鱼地址；Behavior Agent 用真实训练的 GradientBoostingClassifier 替换阈值规则（holdout AUC 0.876 vs 规则基线 0.629；模拟数据 + 熔断阈值把关，避免过拟合结果上线）。AML / Behavior / Phishing agents in parallel — Neo4j GraphRAG for peeling / mixer / layering, Milvus phishing lookup; Behavior Agent GradientBoostingClassifier holdout AUC 0.876 vs rule baseline 0.629, with simulated data + circuit-breaker so an overfit score cannot ship.
• ⚙ 跨项目集成 — 与 RiskGuard 的 XGBoost 提现风险模型真实 gRPC/HTTP 互调；与自研 TradePilot 钱包预检做过联调验证，不计入 DigiPlus 生产。Live gRPC/HTTP to the RiskGuard XGBoost withdraw model; TradePilot wallet precheck was a side-project integration, not DigiPlus production.
• ✅ 工程化与 CI — go build/vet/test 全绿，GitHub Actions 三 job 流水线。Flink 实时特征管道 / Solana Watcher 在文档里如实标「规划中」，不冒充已完成。CI green (build/vet/test), 3-job GitHub Actions. Flink feature pipeline / Solana Watcher labeled planned — not described as shipped.

项目背景Background

面向信息套利 / 对冲 / 合规审计，把预测市场、情报订阅、人才发现与数据市场打通：LMSR AMM + 5 代理辩论 + 分层 Alpha 信号 + Polygon/USDC/UMA 结算 + Stripe。订阅 ARR $3.1M / 首年 ROI 13× 为业务目标，不是已实现收入。

For information arbitrage, funds, and compliance: prediction markets, intel subscriptions, talent discovery, and a data marketplace — LMSR AMM + 5-agent debate + tiered alpha + Polygon/USDC/UMA + Stripe. $3.1M ARR / 13× year-1 ROI are business targets, not booked revenue.

技术架构Architecture

规模：19 后端模块 / 19 前端 Panel / 75+ API / 6 组导航。核心模块：alpha_engine / billing (Stripe) / data_marketplace / debate / market_resolver (UMA) / news_ingest / onchain (Polygon+USDC) / sso (Azure AD/Okta) / talent_scorer / webhook (Slack/Discord)。

Scale: 19 backend modules / 19 panels / 75+ APIs / 6 nav groups. Core: alpha_engine, Stripe billing, data_marketplace, debate, UMA market_resolver, news_ingest, Polygon+USDC onchain, Azure AD/Okta SSO, talent_scorer, Slack/Discord webhook.

职责范围What I owned

• 🏦 LMSR AMM — Hanson 对数市场评分（b 参数连续报价），base_spread 0.5–1.5% 动态调整（volatility_penalty 接近 0/1 时 +0–1%），Sharpe / VaR，流动性成本下降 60%。Hanson LMSR (continuous b-curve quotes), base_spread 0.5–1.5% (volatility_penalty +0–1% near 0/1), Sharpe / VaR, liquidity cost −60%.
• 🧠 5-agent debate + conformal calibration — 3 LLM 路由 + 启发式 fallback，信号置信度 +25%。3-LLM router + heuristic fallback, signal confidence +25%.
• 💰 Alpha Signals — Free EV≥8% / Pro EV≥2% / Enterprise EV≥0.5% · Free 层 Teaser 字段 nulled。Free-tier teaser fields nulled.
• 🚀 AI market creation — 热点新闻到市场上线 ~8 秒，Resolution Agent 自动结算。Hot news to listed market in ~8s; Resolution Agent auto-settles.
• ⛓ Onchain — Polygon + USDC + UMA · 智能合约结算 / dry-run 防误操作。contract settlement / dry-run guards.
• 📡 SSO + Stripe + Webhook — Azure AD/Okta OIDC + Stripe 订阅生命周期 · Slack/Discord 15 条 webhook 事件。Slack/Discord, 15 webhook events.
• 💹 Data Marketplace — 时序数据分层定价卖给机构客户（目标首年 ARR $1.2M，非已实现）。tiered time-series sold to institutions (year-1 ARR $1.2M is a target, not booked).
• 🎨 统一玻璃态 UI，跨多个内部产品复用同一套皮肤。Shared glass UI reused across internal products.
• 📊 从生产数仓多 BU 联合视图抽取真实交易与行为样本做演示与验收。Live trade and behavior samples from production warehouse union views for demos and acceptance.

项目背景Background

传统体育数据平台的痛点：人工定价慢、单一源比对、新闻滞后、套利空间难识别、推荐千篇一律。AlphaOdds 做成「ML 集成 + LLM 代理辩论 + 多源套利 + 时序 + 新闻聚合 + 反作弊 + Live CV」一体化决策系统。margin 7%→9–11% / 留存 +30% / VIP 转化 +20% / 风险拦截 +95% / 首年 ROI 13.3× 为产品目标；下方回测与验收数字为已测结果。

Legacy sports-data pain: slow manual pricing, single-source comparison, news lag, missed arb, one-size recs. AlphaOdds unifies ML ensemble + LLM agent debate + multi-source arb + time-series + news + anti-cheat + live CV. Margin 7%→9–11% / retention +30% / VIP +20% / risk intercept +95% / 13.3× year-1 ROI are product targets; backtest and acceptance numbers below are measured.

技术架构Architecture

4 层：Next.js 前端 / FastAPI 核心 / LangGraph + ML 集群 / 样本切片 + 生产数仓实时查询（Kafka 规划中）。规模：19 后端模块 / 10 前端 Panel / 83 API / 5 LLM 代理 / 8 数据源监控。

Four layers: Next.js / FastAPI / LangGraph + ML cluster / sample slices + live warehouse queries (Kafka planned). Scale: 19 backend modules / 10 panels / 83 APIs / 5 LLM agents / 8 source monitors.

我个人负责的 9 大模块Nine modules I owned

• 🧮 ML Predictive Models — 5 模型加权（XGBoost 0.28 / LightGBM 0.26 / LSTM 0.22 / Transformer 0.14 / GNN 0.10），8 组 60 维特征（近期状态 / 伤停 / 天气 / 主客场 / H2H / xG·PPDA / 教练战术 / 社媒），胜负+得分区间+让分 7 档+Bivariate Poisson Top10，Conformal 95% CI + SHAP，Backtest 5 年准确率 54.2% / ROI 8.4% / Sharpe 1.32 / Brier 0.198 / AUC 0.748。Weighted ensemble (XGBoost 0.28 / LightGBM 0.26 / LSTM 0.22 / Transformer 0.14 / GNN 0.10), 8 groups / 60-dim (form / injuries / weather / home-away / H2H / xG·PPDA / tactics / social), win-draw-loss + score band + 7-line handicap + Bivariate Poisson Top10, conformal 95% CI + SHAP. 5-year backtest: 54.2% / 8.4% ROI / Sharpe 1.32 / Brier 0.198 / AUC 0.748.
• ⚡ Arbitrage Monitor — 8 家数据源，Σ(1/best_price)<1 + Kelly，ARIMA(1,1,1)+LSTM 7 步预测，GPT-4o + Grok 交叉验证新闻；敞口仪表板 $5.27M open / 180ms lag / 24h 1240 条新闻。8 sources, Σ(1/best_price)<1 + Kelly, ARIMA(1,1,1)+LSTM 7-step, GPT-4o + Grok news; $5.27M open exposure / 180ms lag / 1240 news/24h.
• 🧠 5-Agent Debate Engine — LangGraph 状态机驱动 Research / Stats / Odds / Risk / Synthesis 5 异构 LLM（Claude Sonnet 4.5 / GPT-4o / DeepSeek-V3 / Grok 4 / Claude Opus）tit-for-tat 3 轮辩论，权重 Stats 0.24 + Research 0.22 + Synthesis 0.20 + Odds 0.20 + Risk 0.14 + Conformal 校准，3 LLM 路由 + 启发式 fallback。LangGraph state machine: Research / Stats / Odds / Risk / Synthesis on Claude Sonnet 4.5 / GPT-4o / DeepSeek-V3 / Grok 4 / Claude Opus; 3-round tit-for-tat; weights 0.24/0.22/0.20/0.20/0.14 + conformal; 3-LLM router + heuristic fallback.
• 💹 Dynamic Pricing Engine — ML fair value + AMM 自动定价，异常资金流 / 价格突变 / liability / CLV，风险状态 GREEN/YELLOW/RED，24 场 × 多市场 ~150 quote / 5 项运动。ML fair value + AMM, flow/spike/liability/CLV, GREEN/YELLOW/RED, 24 events × ~150 quotes / 5 sports.
• 🛡 Risk Engine — High-Value User（活跃度≥80）/ Arbitrage（1/price_a+1/price_b<0.985）/ Incentive Abuse / Collusion（IP/设备/NetworkX）四类反作弊 + v2.3 User Wellness（LSTM+规则，0–100，7 类信号），6 类干预（auto_limit / cooling / self_assessment / helpline / manual_review / pause）。Four anti-cheat classes (activity≥80, arb 1/a+1/b<0.985, incentive abuse, IP/device/NetworkX collusion) + v2.3 User Wellness (LSTM+rules, 0–100, 7 signals) and 6 interventions (limit / cooling / assessment / helpline / review / pause).
• 🎥 Live CV Pro — MediaPipe Pose + YOLO v8 球员追踪（坐标/速度/跑动/pressing），StatsBomb-style xG + 20×14 热力图；xG 偏差>3pp 自动调价 2–5%；近 10 min xG 激增>0.3 触发 Momentum Shift。MediaPipe Pose + YOLO v8 tracking (coords/speed/distance/pressing), live xG + 20×14 heatmaps; >3pp xG miss auto-reprice 2–5%; 10-min xG surge >0.3 fires momentum shift.
• 🎯 RecSys Pro — DeepFM 16-dim embedding + FM cross + 运动偏好 + tier 加权（Diamond ×1.1 / Gold ×1.05 / Silver ×1.0 / Bronze ×0.95），Half-Kelly（bankroll + 风险偏好 0.2–0.5）+ 4-leg 组合（combined_price × ai_prob > 1）。DeepFM 16-dim + FM cross + sport bias + tier weights (Diamond ×1.1 / Gold ×1.05 / Silver ×1.0 / Bronze ×0.95), Half-Kelly (bankroll + 0.2–0.5 risk) + 4-leg parlays (combined_price × ai_prob > 1).
• 📑 AI Copilot Pro — Bull/Bear/Data/Judge 四段研报 <2 秒，5 语言 zh/en/tl/id/vi，Azure Xiaoxiao/Yunyang + ElevenLabs Adam TTS。Bull/Bear/Data/Judge briefs in <2s, 5 languages zh/en/tl/id/vi, Azure Xiaoxiao/Yunyang + ElevenLabs Adam TTS.
• 📥 Real Data Loader — 生产数仓抽样交易与台账接入 10 个 Panel（不含账号明文）。warehouse sample trades and ledger across 10 panels (no account identifiers).

技术亮点 / 难点Hard parts

• 5 模型加权融合后归一化保证 Σprob=1.0，SHAP 贡献分解出每组特征的方向与自然语言叙述。
• Weighted fusion renormalized so Σprob=1.0; SHAP explains each feature group in natural language.
• 用 LLM 把新闻处理延迟从 5–15min 压到 <60s，自动量化 impact_pp 并触发调价。
• LLM news path cut latency from 5–15 min to <60s, auto-quantified impact_pp and triggered repricing.
• 真实 SABA2 实测 margin 9.42% vs mock 模型差 0.22pp，B2B 客户验收通过率 60%→92%。
• Live SABA2 margin 9.42% vs mock delta 0.22pp; B2B acceptance 60%→92%.

项目背景Background

DigiPlus 大数据团队基础建设角色 — 为上述生产产品提供稳定可靠的数据服务，支撑离线 + 实时全栈链路（数仓分层 ODS/DWD/DWS/ADS、指标口径、SLA）。

Platform role for DigiPlus big-data: reliable batch + realtime services for the production products above (ODS/DWD/DWS/ADS, metric governance, SLAs).

职责范围What I owned

• 🔧 离线 + 实时统一平台 — MaxCompute 批处理 + Spark SQL / Hive + Iceberg 湖仓 + Realtime Compute for Flink + DataWorks + Hologres；数仓分层 ODS/DWD/DWS/ADS，ClickHouse / SelectDB(Doris) 查询加速。Unified platform — MaxCompute batch + Spark SQL / Hive + Iceberg lakehouse + Flink realtime + DataWorks + Hologres; ODS/DWD/DWS/ADS; ClickHouse / SelectDB (Doris) serving.
• ⚡ Flink 生产运维 SOP — Backpressure（busyTimeMsPerSecond / idleTimeMsPerSecond）、Checkpoint/Savepoint（RETAIN_ON_CANCELLATION + 5 retained + incremental）、Watermark/Window incremental aggregation、State Backend（HashMap vs RocksDB + SPINNING_DISK_OPTIMIZED_HIGH_MEM）、keyBy 加盐 + LocalKeyByPartition + 两阶段聚合、KafkaSource Offset 失效改 timestamp 启动、State Processor API 把几十 GB savepoint 裁到几 GB、异常恢复 SOP（--allowNonRestoredState / partition.discovery.interval.ms）。Flink SOP — backpressure (busy/idle ms), checkpoint/savepoint (RETAIN_ON_CANCELLATION, 5 retained, incremental), incremental window agg, HashMap vs RocksDB, salted keyBy + two-stage agg, KafkaSource timestamp startup on bad offsets, State Processor API shrinking tens-of-GB savepoints, recovery SOP (--allowNonRestoredState / partition.discovery.interval.ms).
• 📨 DataHub + Kafka — P99 <200ms, 日吞吐 7+ 亿事件700M+ events/day.
• 📊 MaxCompute SQL 性能调优规范（公司级）— 多个 ADS 任务从 4h+ 优化到 <30 min。Company-wide MaxCompute SQL playbook — ADS jobs 4h+ → <30 min.
• ☁ DataWorks 任务编排 — 任务依赖、口径治理、数据质量与 SLA；指标中台意识，避免口径漂移。DataWorks orchestration — deps, metric governance, data quality, SLAs; shared metric definitions.

项目背景Background

双项目联动：StrategyForge 负责策略回测/参数优化/模拟盘，TradePilot 负责真实下单执行 + LLM 多代理客服/分析 + HITL 风控确认。验证「策略生成」到「真实执行」全链路用真实 Spring Boot 微服务 + MCP 协议工具打通。

Paired systems: StrategyForge backtests/optimizes/paper-trades; TradePilot executes live + LLM multi-agent ops + HITL risk. End-to-end via Spring Boot microservices + MCP tools.

职责范围What I owned

• ☕ Java 21 + Spring Cloud Gateway 微服务集群 — TradePilot 8 服务（gateway / order / position / wallet / market-data / risk-guard / conversation / agent-orchestrator）/ StrategyForge 7 服务，order-svc / position-svc 真实 JPA 持久化（此前 position-svc 无存储层、每次现造假数据）。Java 21 + Spring Cloud Gateway — TradePilot 8 services (gateway / order / position / wallet / market-data / risk-guard / conversation / agent-orchestrator) / StrategyForge 7; order-svc / position-svc on JPA (position-svc previously had no store and invented data per request).
• 🤖 MCP 工具注册 + HITL 风控确认 — 反射式注册 13 个工具，下单/风控类保留纯 metadata 防止绕过人工确认；另用官方 MCP Python SDK 实现真协议 server/client。13 MCP tools via reflection; order/risk stay metadata-only so HITL cannot be skipped; official MCP Python SDK server/client.
• 📈 真实回测引擎 — bar-by-bar 回测，golden-file 测试锁定精确 Sharpe/Sortino；修复过年化收益符号 bug 与 Spring 构造函数歧义 bug。Bar-by-bar backtester, golden-file Sharpe/Sortino; fixed annualized-return sign bug and Spring ctor ambiguity.
• 🔐 API Key 网关鉴权 — 排查发现两个网关此前完全没有鉴权，新增 GlobalFilter 中间件，现场验证 401/放行/健康检查三种情况。API-key GlobalFilter after finding both gateways had no auth; verified 401 / allow / health.

项目背景Background

体育交易场景下单侧对撮合引擎，验证 mutex 保护 vs ring buffer 无锁队列在真实高并发下的表现差异，并堵住钱包/订单接口此前 0 测试覆盖带来的资金安全隐患。

Sports-trading matching engine: mutex vs lock-free ring buffer under real concurrency, plus closing a zero-coverage hole on wallet/order money paths.

职责范围What I owned

• ⚡ Fiber + gnet 双协议网关 — 5K 并发连接头对头 benchmark，如实记录 ring buffer 只有 8+ goroutine 时才反超 mutex 这个不完全符合直觉的真实结果。Fiber + gnet dual gateway — 5K-conn benchmark; ring buffer only beat mutex at 8+ goroutines.
• 💰 撮合引擎 + 钱包/订单模块 — 价格-时间优先撮合，补齐 0 覆盖的下单/冻结/释放资金链路 33 个测试。Price-time matching; 33 tests covering place/freeze/release that previously had zero coverage.
• 🔒 IDOR 越权漏洞修复 — 排查发现钱包/订单接口直接信任客户端 user_id，新增 API Key 鉴权中间件 + 越权回归测试。IDOR fix — wallet/order trusted client user_id; API-key middleware + regression tests.

项目背景Background

对标 Portkey / Helicone / LiteLLM Proxy 的自研 Go AI 网关，挡在业务代码和 OpenAI / Anthropic 等 LLM 供应商之间，业务方只需把 base_url 指向网关、协议保持 OpenAI 兼容格式，统一收敛限流熔断、语义缓存、成本追踪与多供应商故障转移。

A Go AI gateway modeled on Portkey / Helicone / LiteLLM Proxy, sitting between business code and OpenAI / Anthropic — one policy for rate limiting, circuit breaking, semantic caching, cost tracking and failover behind an OpenAI-compatible API.

职责范围What I owned

• 🚦 令牌桶限流 + 三态熔断器（closed/open/half-open）均做内存/Redis 双后端；2000 并发压容量 1000 的令牌桶实测精确放行 1000 次；两个网关实例 + 隔离 Redis 实测限流配额、精确缓存命中、熔断状态三项均跨实例共享（Lua 脚本保证原子性）。Token-bucket limiter + 3-state circuit breaker, both memory/Redis dual-backend; 2000 concurrent requests against a 1000-capacity bucket allow exactly 1000; two gateway instances + isolated Redis verified rate-limit quota, exact-cache hits, and breaker state all shared across instances (Lua for atomicity).
• 💾 两级缓存 — 精确缓存（SHA-256 请求指纹 + TTL）+ 语义缓存（向量余弦相似度）；语义缓存从词袋 hash 近似升级为接入本地 sentence-transformers sidecar 的真实 embedding，"取消订单" vs "退订" 相似度从 ≈0 提升到 0.68，阈值随之从 0.92 调到 0.6（附实测数据）。Two-tier cache — exact (SHA-256 fingerprint + TTL) and semantic (cosine similarity); upgraded the semantic layer from bag-of-words hashing to a real embedding via a local sentence-transformers sidecar, lifting "cancel order" vs "unsubscribe" similarity from ≈0 to 0.68 and retuning the threshold from 0.92 to 0.6 with measured data.
• 📡 SSE 流式透传 + OpenAI / Anthropic 双供应商真实 HTTP 客户端（各自鉴权协议差异真实实现），按熔断器状态自动 failover；OpenTelemetry 全链路 span 覆盖限流 / 缓存 / 供应商调用 / 记账各阶段，按 team 维度的成本账本。SSE passthrough + real OpenAI / Anthropic HTTP clients (each provider's own auth scheme), automatic failover by breaker state; OpenTelemetry spans across rate-limit / cache / provider-call / billing stages, plus a per-team cost ledger.
• 🔀 Redis 迁移设计成一个开关 — 默认零依赖纯内存单机模式，GOGATE_REDIS_ADDR 非空时限流器 / 两级缓存 / 熔断器状态整体切到 Redis 后端，多实例共享状态，Redis 抖动时 fail-open 而不是整体拒绝服务。Redis migration is a single switch — default zero-dependency in-memory mode; setting GOGATE_REDIS_ADDR moves limiter / cache / breaker state to Redis for multi-instance sharing, fail-open on Redis hiccups instead of denying service.

项目背景Background

Go 生态里对标 LangGraph 的编排原语基本空白——AgentFlow 提供 Graph + State + Node + Executor 这套最小必要抽象：顺序执行、并行 fan-out/fan-in、按运行时 State 条件分支、超时重试、HITL 人工确认中断-恢复，核心库不依赖任何 LLM SDK，可以和自研 GoGate 网关组合使用。

Fills the LangGraph-shaped gap in Go — Graph + State + Node + Executor primitives for sequential execution, parallel fan-out/fan-in, State-driven conditional branching, timeout/retry, and HITL interrupt/resume, with zero LLM-SDK dependency in the core; pairs with GoGate as the gateway layer underneath.

职责范围What I owned

• 🧩 图执行引擎 — 并发安全 State（泛型 Get[T]/GetOr[T]），条件分支就是 Node.Run 的返回值而非独立抽象；Parallel 用真实 goroutine + WaitGroup 并发，测试用可控 sleep 时长断言总耗时低于顺序执行下限，证明真并发而非伪装。Graph execution engine — concurrency-safe State (generic Get[T]/GetOr[T]); branching is just a Node.Run return value, not a separate abstraction; Parallel runs real goroutines + WaitGroup, with tests asserting wall time below the sequential lower bound to prove genuine concurrency.
• ⏸ HITL 人工确认中断/恢复 — WithInterruptBefore 在敏感节点（如放款）前真正暂停：调用栈已完全退出、HTTP 请求已正常返回，*Paused 哨兵类型带着待恢复节点名；Resume(ctx, fromNode, state) 从原地续跑不重跑历史节点；State 实现 MarshalJSON/UnmarshalJSON，测试验证序列化后丢弃原指针、反序列化出新对象仍可正确 Resume。HITL interrupt/resume — WithInterruptBefore genuinely pauses before sensitive nodes (e.g. releasing a refund): the call stack has fully returned and the HTTP request already completed; a *Paused sentinel carries the pending node name, and Resume(ctx, fromNode, state) continues in place without replaying prior nodes; State implements MarshalJSON/UnmarshalJSON, with a test proving a serialize-discard-deserialize round trip still resumes correctly.
• 🎫 电商客服工单分诊示例 — Classifier → Retriever → Responder → Reviewer → RefundApproval 六节点图，退款类工单在文案审核通过后仍需进入独立的人工放款审批关卡，cmd/agentflow-server 暴露 /api/run 与 /api/runs/{id}/resume 两个真实端点。E-commerce support-triage example — a six-node graph (Classifier → Retriever → Responder → Reviewer → RefundApproval); refund tickets still hit a separate human approval gate even after the reply passes review, with cmd/agentflow-server exposing real /api/run and /api/runs/{id}/resume endpoints.
• 🔍 向量检索 — 接入真实 Qdrant（Cosine 距离 + HNSW 索引 + ANN top-K），Qdrant 不可达时自动降级为关键词匹配兜底；中文语料没有空格分词，改用汉字 unigram/bigram，并现场排查出一次哈希碰撞导致的误召回（无关查询相似度反而更高），加大哈希维度后修复。Vector retrieval — wired to real Qdrant (cosine distance + HNSW + ANN top-K), auto-degrading to keyword matching when Qdrant is unreachable; switched Chinese tokenization to character unigrams/bigrams (no whitespace boundaries), and diagnosed a live hash-collision bug where an unrelated query scored higher than relevant ones, fixed by widening the hash dimension.

项目背景Background

面向概率型游戏机制（如老虎机等随机结果类玩法）的数学模型工作台——闭式 RTP 分解 + 全窗口穷举交叉验证 + 大规模仿真标定，落地成 Python 后端 + FastAPI 服务 + Next.js 策划工作台，覆盖游戏后端 API 设计、确定性 RNG、结果可复现审计与发布门禁。

A math workbench for probability-driven game mechanics (such as slot-style random-outcome games) — closed-form RTP breakdown, exhaustive cross-validation, and large-scale simulation calibration, shipped as a Python backend + FastAPI service + Next.js planning UI covering game-backend API design, deterministic RNG, replayable audit, and release gating.

职责范围What I owned

• 🎰 OutcomeEngine 状态机（spin → RoundResult）+ 确定性 RNG，配合 seed 复现工具做结果审计，杜绝"线上分数对不上账"。OutcomeEngine state machine (spin → RoundResult) + deterministic RNG, paired with a seed-replay tool for result audit — no "the payout doesn't match the log" gaps.
• 🧮 RTP/命中率/Feature 触发率闭式期望推导与全窗口穷举交叉验证双路径互证，不是拿仿真估算冒充精确解；不同 RTP 档位（96%/94%）权重逐位相同、只有赔付缩放不同，用测试锁定。Closed-form expectation for RTP / hit rate / feature-trigger rate cross-verified against exhaustive full-window enumeration — not a simulation estimate dressed up as exact; two RTP tiers (96%/94%) share identical weights with only payout scaling different, locked by tests.
• 🔌 FastAPI 暴露 /v1/spin /v1/theory /v1/sim /audit/replay，含配置结构校验与发布门禁（spec_fail_reasons）；77 条 pytest 用例（含 2 条全窗口穷举回归）。FastAPI exposes /v1/spin /v1/theory /v1/sim /audit/replay with config validation and a release gate (spec_fail_reasons); 77 pytest cases including 2 exhaustive full-window regressions.
• 🔍 两轮自我代码评审（docs/REVIEW.md）：发现并修复幂等、性能（19.4s→0.75s）与配置校验类 bug，记录发现-修复-验证全过程。Two self-review passes (docs/REVIEW.md): found and fixed idempotency, performance (19.4s→0.75s), and config-validation bugs, with the full find-fix-verify trail documented.

项目背景Background

补 RiskGuard（事后分析）和 ChainPayGraph（链上闸口）都不覆盖的一环：法币订单进通道前做 ALLOW / CHALLENGE / REJECT，通道 webhook 再对账。虚构商户与通道、本地造单，不接真实银行卡、不做链上、不做撮合。

Fills the gap neither RiskGuard (post-settlement) nor ChainPayGraph (on-chain gate) covers: ALLOW / CHALLENGE / REJECT before a fiat order hits a channel, then reconcile webhooks. Fictional merchants and channels, locally generated orders — no real cards, no chain, no matching.

职责范围What I owned

• ☕ Java 17 + Spring Boot 3.3 ledger-api — Flyway 20 张业务表；支付下单、四条硬规则、对账 run、成功率指标、人工 override；决策权只在 Rule ∪ Human，没有任何路径让 LLM 写 risk_decision.decided_by 或改订单状态。Java 17 + Spring Boot 3.3 ledger-api — Flyway 20 business tables; payments, 4 hard rules, recon runs, success-rate metrics, human override; decision authority is Rule ∪ Human only — no path for an LLM to write risk_decision.decided_by or flip order status.
• 🔐 人工 override 原先信任请求体里的 actor 字符串、可伪造审批人并翻转 REJECT→ALLOW；改为 X-LDR-Reviewer-Key 映射 reviewer，audit actor 只来自已认证密钥，缺 key 返回 401。Human override used to trust an actor string in the body (anyone could forge a reviewer and flip REJECT→ALLOW); now X-LDR-Reviewer-Key maps to a reviewer and the audit actor comes only from the authenticated key (401 if missing).
• 🤖 可选 ldr-ai（FastAPI）只润色规则已给出的文案，超时/无模型时回落规则原文、行为不变；mcp-ldr 暴露 3 个只读 MCP 工具；Next.js 值班台覆盖订单 / 对账。Optional ldr-ai (FastAPI) only polishes wording the rule already produced — timeout / no model falls back to the rule text with no behavior change; mcp-ldr exposes 3 read-only MCP tools; Next.js duty console covers orders / recon.

项目背景Background

用户风险行为管理工作台：Soft Circuit 决策树 + 异常消费/连续加大投入检测 + 合规模板，看板可接 MaxCompute。AlphaOdds 里有一份轻量同类模块，细节以本仓库为准。不计入 DigiPlus 生产。

User risk-behavior workbench: Soft Circuit decision tree, abnormal-spend detection, and compliance templates, with optional MaxCompute dashboards. AlphaOdds carries a lighter version of the same module; this repo is the source of truth. Not DigiPlus production.

职责范围What I owned

• 🛡 Soft Circuit 静态树 + 异常消费检测（会话窗口近似）已在真实 ODPS 用户/会话表上跑；六级干预动作库与合规模板。Soft Circuit static tree + abnormal-spend detection (session-window approximation) run on real ODPS user/session tables; six-level intervention catalog and compliance templates.
• 🧠 Behavioral AI 是真训的 sklearn GBDT（生物识别代理 + 点击流 + 词典情绪），权重在仓库内；标签和遥测目前是 mock，AUC 来自这份合成训练，不拿生产标签吹。Behavioral AI is a real sklearn GBDT (biometric proxies + clickstream + lexicon sentiment) with weights in-repo; labels and telemetry are mock, so AUC is from that synthetic train — not production labels.
• 🔒 查真实 login_name / user_id 的接口要求 X-API-Key，挡住匿名枚举画像；SQL 注入用白名单校验。PLE 多任务只做架构展示，推理是哈希模拟，没有权重文件。Endpoints that look up real login_name / user_id require X-API-Key to block anonymous profiling; SQL injection is allowlist-checked. PLE multi-task is architecture-only — hash-simulated inference, no weight files.

项目背景Background

三条可独立跑、又可互相调用的增长工作台：LTVForge 看用户价值，CausalMix 做因果/Uplift，BonusForge 用 OPE 判促销策略能不能上线。看板可选接 MaxCompute；TFT / DeepHit / PPO 等深度模型是架构展示，不拿未训练权重当生产模型。

Three workbenches that run alone or call each other: LTVForge for user value, CausalMix for causal/uplift, BonusForge for promo OPE go-live gates. Optional MaxCompute; TFT / DeepHit / PPO-class deep models are architecture demos — no untrained weights sold as production models.

职责范围What I owned

• 📈 LTVForge — Cohort / 鲸鱼帕累托 / 冷启动路径；真正训过的是 LightGBM baseline（ODPS cohort → holdout MAE）。TFT / DeepHit 只保留结构图与接口，没有 torch 权重。LTVForge — cohorts / whale Pareto / cold-start paths; the trained model is a LightGBM baseline (ODPS cohort → holdout MAE). TFT / DeepHit stay as architecture + reserved APIs, no torch weights.
• 🧪 CausalMix — 27 个 API 里约 5 个接真实 ODPS，其余为教学/合成。真链路是渠道映射 + Double ML（register_cnt → GGR，5-fold cross-fitting）；MMM / 预算 What-If 主体仍是 mock。Uplift 分群 HTTP 给 BonusForge 约束 Do-Not-Disturb 用户不发 Bonus。CausalMix — ~5 of 27 APIs hit real ODPS, the rest teaching/synthetic. The live path is channel mapping + Double ML (register_cnt → GGR, 5-fold cross-fitting); MMM / budget What-If stay mock. Uplift-segment HTTP lets BonusForge withhold bonus from Do-Not-Disturb users.
• 🎁 BonusForge — IPS / SNIPS / DR 是真公式（行为策略来自 ODPS 促销曝光，评估策略跑在分层抽样用户上）；PPO / Double DQN 是规则+扰动的 Policy Simulator，不是训过的神经网络。CausalMix 不可达时 fail-open 并在响应里标明约束未生效。BonusForge — IPS / SNIPS / DR are real formulas (behavior policy from ODPS promo exposure, evaluation on stratified users); PPO / Double DQN is a rule+noise policy simulator, not a trained net. CausalMix downtime fail-opens and the response flags that the constraint did not apply.