为京东物流加拿大的下一步而来。协同仓已经在跑，真正的难题是把仓填满、把方案沉淀成能复制的打法，并算清楚哪一笔量该转自营。这正是我在香港做过的事：自营柜机与 800 多家合作门店跑在同一套订单系统上，网点从一千做到两千多，硬件资本没有再加。
Built for JD Logistics Canada's next step. The co-managed warehouses are already running; the hard part now is filling them, turning solutions into a playbook others can run, and knowing which volume justifies going self-operated. That is exactly the work I've done before: self-operated lockers and 800+ partner storefronts on one order system, coverage from 1,000 to over 2,000 points with no additional hardware capital.
这个岗位真正的产出不是一份方案，是一支能不断产出方案的团队。第一年可能是一个人扛，但如果三年后加拿大还是靠一个人写方案，那这个位置就没建成。这件事我做过：顺丰五年，直接带 20 人，BU 从 3 个人做到四五十人。我的习惯是把每一单都沉淀成别人能用的东西：SOP、报价模板、KPI 骨架、实施 checklist。方案做完、人走了还在，那才叫打法。
底下的功夫也不缺：在蒙特利尔两个多月从零建仓上线；在香港从零架构最大的最后一公里生态（OMS、TMS、柜机操作系统、微型仓 WMS、IoT 硬件与承运商 API），做到年处理 1,100 万件包裹。做法一贯是数据为先（Tableau/Metabase KPI 体系直接给出行动建议，而不止于看板）、流程为先（先重设拣货路径与工作流，再谈自动化）。
What this role has to produce is not a solution — it is a team that can keep producing them. Year one may well be one person carrying it, but if Canada still depends on a single person to write solutions three years from now, the seat was never really built. I have done this before: five years at SF Express with 20 direct reports, in a BU that grew from three people to forty or fifty. My habit is to turn every engagement into something others can run: SOPs, rate-card templates, KPI skeletons, implementation checklists. A solution that outlives the person who wrote it is what makes it a playbook.
The hands-on depth is there too: a greenfield warehouse launched in Montréal in a little over two months, and Hong Kong's largest last-mile ecosystem (OMS, TMS, locker OS, micro-warehouse WMS, IoT hardware and carrier APIs) architected from zero to 11 million parcels a year. The approach is consistent: data-first (Tableau/Metabase KPI systems that prescribe actions, not just dashboards), process-first (pick-path and workflow redesign before automation).
顺丰的柜网铺到一千个点之后，再加一个点就是纯资本投入。我去谈了便利店合作，八百多家店接进来。关键不是签约，是让合作点位跟自营柜机跑在同一套订单和预约系统上，客户根本感觉不到哪个点是自营、哪个是合作方。覆盖从一千做到两千多，硬件资本没有再加一分。
Once the SF Express locker network hit 1,000 points, every additional point was pure capital. So I negotiated the convenience-store partnership and brought 800+ storefronts online. The hard part wasn't signing them — it was putting partner sites and self-operated lockers on the same order and reservation system, so customers could not tell which was which. Coverage went from 1,000 to over 2,000 points with no additional hardware capital.
合作点位与自营点位共用订单、预约、履约与结算，客户体验没有差别；运营看到的是一张网。系统不统一，协同仓就永远只是外包。Partner and self-operated nodes share one order, reservation, fulfillment and settlement stack. Without that, a co-managed warehouse is just outsourcing.
SLA 分级、异常率看板、按节点绩效数据结算、OS&D 责任划分写进合同。管得住才敢扩，这是当年从一千扩到两千的前提。Tiered SLAs, exception-rate dashboards, settlement driven by per-node performance data, OS&D liability written into the contract. You only scale what you can govern.
三笔账：协同仓单均成本乘以量；自营的固定成本（租金、设备、人力）加变动成本；两条线的交叉点落在哪个量级。给出 IRR 与回收期，并把触发条件写死：月单量稳定超过多少、合同期客户占比超过多少。Three models: per-unit co-warehouse cost × volume; self-operated fixed cost (rent, equipment, labor) plus variable; and where the two curves cross. Then IRR, payback period, and explicit triggers.
量没起来就自营，固定成本会压死你；量起来了还不自营，单均降不下来、服务受制于人。这个判断只能靠模型，不能靠感觉。Go self-operated too early and fixed costs crush you; too late and unit costs stay high while service stays outside your control. This call needs a model, not instinct.
先把边界说清楚：纯大件仓我没有直接运营过。但方案方法论是相通的，而大件恰恰是最不能凭经验拍板的品类，一个承重或通道宽度算错，改造成本比省下来的多得多。所以我的顺序是先做三件事：数据画像、现场评估、安全评估，然后才画图。
Boundary first: I have not run a dedicated bulky-goods warehouse. The solution methodology carries over, but bulky is the category where guessing is most expensive — get rack capacity or aisle width wrong and the retrofit costs more than the savings. So the order is: data profile, site assessment, safety assessment. Drawings come after.
材积与重量分布、是否成套、破损敏感度、送装要求（是否上楼、拆包、安装）、退货率与翻新流程。画像不清楚，后面每一步都是猜。Cube and weight distribution, set-based SKUs, damage sensitivity, installation requirements (stairs, unpacking, assembly), return rate and refurbishment flow.
货架承重与高位限制对比地堆与横梁架、通道宽度按设备转弯半径反推、叉车与夹抱选型、双人搬运安全规程、下重上轻并按出货频率排位。Rack capacity and height limits vs floor stacking and beam racking, aisle width derived from equipment turning radius, forklift and clamp selection, two-person handling protocols, heavy-low placement ordered by pick frequency.
大件不走拣选线，走整件直发区；装车按配送路线倒装，后送的先装。这一条做错，末端每一趟都在返工。Bulky bypasses the pick line and moves through a full-case direct-ship zone; trucks are loaded in reverse route order, last stop loaded first. Get this wrong and every route pays for it.
两人组配送、预约制、白手套（送装、拆包、包装回收），尾程是成本大头。计费按材积与重量分档，超规附加，破损险单列。Two-person crews, appointment scheduling, white-glove service (delivery plus install, unpacking, packaging removal). Final mile dominates cost. Price in cube/weight tiers, with oversize surcharges and damage cover priced separately.
香港最大的最后一公里网络：1,000+ 智能柜、800 家便利店、年处理 1,100 万件包裹——OMS、TMS、硬件与承运商 API 一体化设计。从零到市场第一，5 年。
Hong Kong's largest last-mile network: 1,000+ lockers, 800 convenience stores, 11M parcels/year — OMS, TMS, hardware and carrier APIs designed as one system. Zero to market leader in 5 years.
AI 系统 · 多智能体AI Systems · Multi-agent把 PRD、路线图与 SQL 库变成 AI 可读的结构化上下文，三个角色智能体（产品 / 工程 / 分析）在同一张产品知识图上工作。一个人从零设计并落地。
Turns PRDs, roadmap and the SQL library into structured, agent-readable context — three role-based agents (product / engineering / analytics) working off one product knowledge graph. Designed and built solo.
仓库执行 · 劳动力编排Warehouse Execution · Labor orchestrationFoodsUp WES 产品负责人。排班模块把每周四十几人、三百多条排班收进一套可审计的系统：19 个班次、9 个岗位、14 家劳务公司。另有 KPI 基准报表底座、波次配置、司机装载与拣货异常看板。人工优先阶段最需要的正是这一层。（案例页展开的是其中的任务与排班模块）
Product owner of FoodsUp's WES. The labor module brings 40-odd staff and 300+ weekly shift assignments into one auditable system: 19 shift patterns, 9 job functions, 14 staffing agencies. Plus the KPI baseline reporting layer, wave configuration, driver loading and pick-exception boards. This is the layer a manual-first operation needs most. (The linked case study details the task and scheduling module.)
运输 · AITransportation · AI面向末端配送的 AI 路径优化——聚类 + 约束求解，覆盖两个枢纽、每月 48K 订单。运输侧的算法实践。
AI-orchestrated route optimization for last-mile delivery — clustering + constraint solving over 48K orders/month across two hubs. Algorithms applied to the transportation side.