AI-Driven · Smart Replenishment · Data-Backed Decisions
An AI-powered replenishment decision platform for small and mid-sized e-commerce and retail businesses. Sales, inventory, in-transit and product master data converge into a single fact foundation, with two built-in business pipelines: 2C self-operated weekly replenishment and office-supply omnichannel replenishment. AI agents guide business users through data fetching, formula computation, plan adjustment and replenishment order generation — every suggestion backed by data, formulas and reports, fully traceable.
Core Jikeyun APIs
2C weekly + office omnichannel
General + fetch + compute
What the platform does, distilled into three visual scenario maps — use the buttons to jump to the matching chapter.
Sales, inventory, in-transit and product master data settle into one fact foundation; the knowledge base retrieves with 1536-dimension hybrid vectors so AI truly understands business data.
Explore the data foundation
The general assistant, fetch engineer and compute engineer collaborate around real supply-chain work. Business users describe goals in natural language; agents handle fetching, computing, plan adjustments and order generation.
See how AI works
Fetching, processing, computing, planning and ordering — traceable end to end. Replenishment rules are scripted with tunable coefficients; every suggestion traces back to concrete data and formulas.
See the full pipelineA three-tier architecture (A data detail → B fetch & process → C compute) plus D plans & orders: every replenishment suggestion traces line by line back to its data, formula and adjustment history.
The fact foundation: data from upstream APIs, Excel imports and manual upkeep is persisted under unified primary keys.
Configuration-first: business users define fetch rules that run instantly in the sandbox, producing dimension-aligned detail tables.
Compute scripts consume B-layer output: primary fetch persisted → async BullMQ batch computation → row-level state-machine closure.
Computation results become executable replenishment plans, then split into replenishment orders by caliber.
Business scenarios are freely categorized (no enum binding); new scenarios require no schema changes.
Built on Qeasy's connectivity to 500+ mainstream systems — Kingdee, Yonyou, SAP, Jushuitan, Jikeyun, Mabang and more across ERP, e-commerce and supply chain. The replenishment scenarios showcased here focus on Jikeyun open-platform APIs plus manual ledger adapters.
| Data domain | API | Purpose |
|---|---|---|
| Warehouses | erp.warehouse.get | Master data / multi-warehouse calibers |
| Product categories | erp.goodscate.get | Master data / full replacement |
| Products | erp.storage.goodslist | Master data / products + SKUs |
| Suppliers | erp.vend.get | Master data / turnover dictionary |
| Sales channels | erp.sales.get | Master data / channel dictionary |
| Purchase orders | erp.purch.get | One source of in-transit orders |
| Real-time inventory | erp.stockquantity.get | Available warehouse stock |
Fetch scripts emit dimension-aligned structured detail: server-side pagination with client-side search, filter and sort. Previews and small-data debugging run as synchronous dryRuns; production computation runs on the async BullMQ pipeline.
Built on the platform agent runtime, every write action goes through the existing script + execution chain, with a business-language confirmation before triggers. Agent retrieval relies on a knowledge base with hybrid vector + keyword recall, RRF fusion and MMR re-ranking — no invented abstraction layers.
general-assistant
Master data / knowledge base / system capability queries
Read-only tools; no write operations
fetch-script-engineer
Atomic fetching with FetchScript
One business-language confirmation before a trial run
compute-script-engineer
Batch computation with ComputeScript + replenishment plan templates
One business-language confirmation before submitting a plan
Security and observability are built into the platform baseline, not patched in afterwards.
Layers B/C share one sandbox: dual-layer isolation with child_process + isolated-vm; fetch returns synchronously within 30s, compute runs in async batches.
Results / parameters / row counts / durations are written to a shared audit table — business rows are not materialized, keeping traceability without storage bloat.
BullMQ queues + 200k-row circuit breaker + 500-row batch fan-out; failed rows can be retried at plan level.
Single-tenant today; extension seams such as SupplierTenantBinding are reserved for future business needs.
Primarily the 7 core Jikeyun open-platform APIs (warehouses / products / categories / suppliers / sales channels / purchase orders / real-time inventory); Excel adapters import custom raw tables.
Layer B aligns layer-A details into consumable structures; layer-C compute scripts consume B's output into PENDING rows; BullMQ async batches produce suggestions, which then go through row-level confirmation → order splitting → orders.
Three agents are registered (general assistant + fetch script engineer + compute script engineer); every write action requires one business-language confirmation and goes through the existing script and execution chain.
Yes. The Qeasy data integration platform has accumulated connectivity to 500+ mainstream systems — Kingdee, Yonyou, SAP, Jushuitan, Jikeyun, Mabang and more across ERP, e-commerce and supply chain. Each project confirms scope and implementation around business goals, API contracts, field mapping and data calibers.
Computation results can be traced row by row: dimensions, raw variables, formulas and adjustment records are all persisted; plan recomputation produces versioned batches with parentPlanId lineage preserved.
Register to get started, or contact Guangdong Qingyiyun Software Technology Co., Ltd. to learn about onboarding.
