Research

自动化生成系统 PRD：PostgreSQL 提取、LLM 分类、Vega-Lite 图表到 HTML 报告的全链路设计。含交叉筛选、LLM 分析面板。Automated report generation system PRD: from PostgreSQL extraction, LLM classification, Vega-Lite charts to HTML report. Cross-filtering, LLM analysis panel included.

从 BI 到 Agent：工单分析与交叉对比场景下的三层可视化技术选型。ECharts、Plotly、D3 学习路径 + Agent 驱动的出图工作流。From BI to Agent: a three-layer viz stack for ticket analysis and cross-comparison. ECharts, Plotly, D3 learning paths + Agent-driven chart generation workflow.

5 年持续发货医疗产品的三张图分析范式：组件级 Pareto、分层浴缸曲线、Cohort 累计失效对比。用 ECharts 交互渲染，每张图驱动一个 Action 决策。A three-chart analysis paradigm for a 5-year medical product in continuous shipment: component Pareto, stratified bathtub curve, and cohort cumulative comparison. ECharts interactive rendering, each chart drives one action decision.

Hermes Agent 长期记忆系统深度解析。架构分层、数据流、进程拓扑、配置方案。从 Agent 到 PostgreSQL 的全链路拆解。A deep dive into Hermes Agent's long-term memory system: architecture layers, data flow, process topology, and configuration. Full-chain teardown from Agent to PostgreSQL.

2026 湖南物理类 458 分女生、物化生组合的志愿路径：师范、医学技术、药学与冲稳保策略。Advisory for a Hunan 2026 physics-track 458 score: education, medical technology, pharmacy, and balanced application strategy.

基于官方位次 156,519 和 45 个公办普通组初筛志愿，生成高冲、冲、稳、保四档建议表，并标注录取可能性与专业组调剂风险。A final advisory table based on official rank 156,519 and 45 public-school groups, with tiers, probability bands, and adjustment risks.

深度解析三大核心问题：志愿顺序是否影响录取？冲刺会抬高投档线吗？真的可以拼吗？附投档机制详解与冲稳保梯度实操框架。Deep-dive into three critical questions about parallel volunteer system: ordering priority, score inflation, and reach strategy. With practical tier framework.

1655 个国产独立项目的田野调查。领域分布、商业模式、定价策略与生存状态，一屏看懂独立开发生态。A field survey of 1,655 indie projects from China — domain distribution, business models, pricing, and survival patterns at a glance.

Serviceability 主业方向的方法论纲领：工单数据 → 故障模式枚举 → 故障树图谱的目标、范围与五年路线。The methodological charter for my serviceability day job: work-order data → failure-mode enumeration → fault-tree graphs, with goals and a five-year roadmap.

AI 志愿填报 agent 的交互原型（v0/v1 迭代稿）：调研问卷、评估流、报告生成的完整用户体验链。Interaction prototypes (v0/v1) for an AI gaokao advisory agent — intake survey, assessment flow, and report generation, end to end.

现代化 APS 的完整方法论：闭环四层架构、四层痛点地图（数据/模型/执行/环境）、六大解法（数据回流、时间栅栏、有界重排、混合求解引擎）与五步落地顺序。A complete methodology for modern APS: closed-loop architecture, four-layer pain-point map, six solutions (data feedback, time fences, bounded rescheduling, hybrid solver), and a five-step rollout order.

月中预排产买票、月初实际订单上车的两阶段循环如何 Agent 化：FJSP + 资质矩阵形式化、LLM 翻译官 + OR-Tools 求解器的四层架构、最小扰动重排与历史重考落地路线。How to agentify the two-stage cycle of mid-month pre-scheduling and month-start firm-order filling: FJSP formalization with skill matrix, a four-layer LLM-translator + OR-Tools solver architecture, minimal-disruption rescheduling, and a backtest-driven rollout path.

从「排产经理被迫用 Excel 双录」到 CP-SAT 引擎黑盒的 0→1 全程：三幕故事版、ELI5 动画、需求纪要、引擎设计文档（冻结三档/两段式求解/无解诊断）、框架图与时序图、五个版本的求解 demo（插撤单/故障/装机死局/工作日历）。五类扰动一个机制：冻结区 + 最小扰动重排。The full 0→1 journey from the planner's forced Excel double-entry to a CP-SAT engine black box: three-act story, ELI5 animation, requirement minutes, engine design doc (three-tier freeze / two-phase solving / infeasibility diagnosis), architecture & sequence diagrams, plus five solver demo versions. Five disruption types, one mechanism: freeze zone + minimal-disruption rescheduling.

用 MCP 驱动真实 Excel 进程：COM 自动化层、MCP Server + CLI 双入口架构、会话生命周期、326 个原子操作，以及相对 Copilot 365 的定位差异。Driving the real Excel process via MCP: COM automation layer, dual MCP Server + CLI architecture, session lifecycle, 326 atomic operations, and how it differs from Copilot 365.

Karpathy 的 AI 输出四阶法：用航空维修手册的受控英语约束大模型。附真实 A/B 对照实验——句长降 60%，字数只降 22%，机制是短句化而非砍字数。附适用边界与维修工单场景落点。Karpathy's four-level output ladder: constrain the LLM with aviation-maintenance controlled English. With a real A/B test — sentence length down 60%, word count down only 22%: the mechanism is shortening sentences, not cutting words. Boundaries and a field-service use case included.