Guangdong Qingyiyun Software Technology Co., Ltd. · Enterprise-grade AI Agent in production
The AI Agent-Powered E-commerce Financial Reconciliation Platform
The Qeasy intelligent reconciliation platform upgrades the 'manual exports + Excel cross-checking' finance workflow of large-scale e-commerce businesses into a traceable, schedulable and extensible system: fact-data engineering, a knowledge base, AI Agent e-commerce reconciliation, custom reports and smart financial reconciliation — putting AI to work in real finance scenarios.
Traceable reconciliation lineage (reverse lookup via minimal bridge tables)
Fact-data engineering + knowledge base
An AI-Driven Data Foundation
Facing e-commerce statements in wildly different formats from JD, Douyin, Alipay, Amazon and more, the platform normalizes heterogeneous statements into a standard fact layer with Decimal(20,4) precision using sandboxed JS parsing scripts, then distills platform rules, accounting items and script samples into a reasoning-ready business knowledge base through pgvector + HNSW vector indexing and hybrid retrieval (vector + keyword → RRF → MMR). AI finally 'reads' e-commerce finance instead of staring at piles of Excel files.
AI Agent capabilities + AI Agent e-commerce reconciliation
The AI Agent Matrix, Applied to E-commerce Reconciliation
The platform ships one general assistant (the global chief of staff) plus three domain agents — Bill Parsing Engineer, Reconciliation Script Engineer and Expense Allocation Engineer. Through a five-step standard workflow (requirement dialogue → automatic reconnaissance → script writing → sandbox testing → execution), finance staff drive reconciliation script authoring, testing and execution in natural language. Agents never bypass scripts to mutate business data: every side effect still goes through the existing 'script + JobTask + BullMQ' pipeline — rollbackable and auditable, working in real e-commerce finance engineering scenarios rather than demo-grade toys.
Smart financial reconciliation + custom reports
Smart Reconciliation, AI-Assisted Decisions
Built for real e-commerce finance at the 100k-document scale, the platform runs twin reconciliation plans (revenue + expenses) on an async BullMQ queue with a 7-state state machine, automatically completing the three-way match across platform statements ↔ supply-chain orders ↔ internal accounting, with item details written back automatically. Together with expense aggregation views, raw statement views and custom reports that drill down by platform / store / business order number, reconciliation results become decision-ready financial insight — AI working for real in large-scale e-commerce finance.
From Manual Excel Checks to AI-Driven Reconciliation
E-commerce agencies, brands and finance outsourcing teams wrestle with multi-platform statements every month. Qeasy turns this workflow into an asynchronous, traceable and re-runnable system capability.
Manual exports + row-by-row Excel checks
→
Statements auto-imported and persisted, driven by async tasks, traceable throughout
Every platform's statement format is different
→
Sandboxed JS parsing scripts auto-identify business order numbers and accounting items for any platform, any format
Discrepancy causes hunted down by hand
→
Dual-code reconciliation compares line by line; variance amounts and reasons persisted per row
Shared expenses allocated by gut feel
→
Sandboxed allocation scripts compute automatically and write back to revenue reconciliation lines in real time
Core capabilities
Six Capabilities Covering the Full E-commerce Reconciliation Lifecycle
Fact-Data Engineering
One-click import persists statements from any platform in any format; each platform × statement type gets its own fact table with full raw-row snapshots; finance-grade Decimal precision (amounts (20,4) / unit prices (20,6)); a 5-field minimal bridge table traces any reconciliation result back to the original statement row.
24 fact tables
Raw-row snapshots (rawData)
End-to-end reverse traceability
Smart Financial Reconciliation
Twin reconciliation plans: revenue auto-matches supply-chain orders (dual-code comparison) while expenses are confirmed line by line; a three-tier expense architecture (plan → aggregation → allocation detail) with sandboxed allocation scripts writes shared expenses back to revenue lines automatically, and item details are written back on success.
Twin revenue & expense plans
Sandboxed expense allocation with write-back
State machine, auditable throughout
AI Agent E-commerce Reconciliation
Users describe business needs in natural language and AI Agents close the loop: analyze → confirm → write script → test → launch → verify. Three domain agents — bill parsing, reconciliation and expense allocation — each do their job, dispatched by the general assistant.
Natural-language-driven reconciliation
Automated script generation and testing
Audited execution trail
Knowledge Base
Hybrid recall with pgvector vector search + keyword search (RRF fusion + MMR re-ranking) over 1536-dimension Chinese semantic vectors; help documents and AI knowledge share one source — maintained once, effective everywhere; async LLM evaluation continuously monitors retrieval quality.
Hybrid vector + keyword retrieval
Help docs and AI knowledge from one source
Continuous retrieval-quality evaluation
Custom Reports
Expense aggregation and raw statement views work out of the box, pivoting freely by platform / store / business order number; dual-dimension aggregation by accounting item + accounting unit, with custom parsing and allocation scripts keeping report calibers flexible as the business evolves.
Pivot by platform × store × order
Dual-dimension accounting-item aggregation
Script-defined report calibers
Enterprise-Grade Platform Foundation
BullMQ async queues power high-volume statement processing; a dual-layer sandbox (isolated-vm + dedicated child processes) isolates user scripts; every agent tool call is persisted to audit tables, script version history is rollbackable, and tests never touch the database — financial data stays safe and under control.
BullMQ async task scheduling
Dual-layer sandbox isolation
Full audit trail + version rollback
Reconciliation workflow
A Five-Step Loop — Every Penny Accounted For
1
Statement import
Upload Excel / CSV from any platform; async tasks persist automatically
2
Sandboxed parsing
JS scripts identify business order numbers and accounting items; failures are fixed and re-queued
3
Pivot views
Expense aggregation and raw statements pivoted by platform, store and order number
4
Twin reconciliation
Revenue auto-matched to supply chain, expenses confirmed line by line, variances persisted
5
Allocation write-back
Shared expenses allocated in the sandbox and written back to revenue lines in real time
Three-way matching across platform statements ↔ supply-chain orders ↔ internal accounting — monthly reconciliation with minimal manual intervention.
AI Agents in production
Not a Demo — an Agent Matrix Doing Real Work at Large-Scale E-commerce Companies
Finance staff state requirements in natural language; agents autonomously reconnoiter, write, test and save — humans only confirm once before execution. Three domain agents each do their job, orchestrated by the general assistant.
bill-parse-agent
Bill Parsing Engineer
Creates, edits, tests and triggers parsing scripts; auto-identifies business order numbers and accounting items; validates parsing results.
reconcile-script-agent
Reconciliation Script Engineer
Writes and tests reconciliation scripts, drives revenue plan execution, and compares supply-chain orders line by line with result verification.
expense-allocate-agent
Expense Allocation Engineer
Manages the full lifecycle of allocation scripts, executes shared-expense allocation with write-back to revenue lines, and validates results.
general-assistant
General Assistant · Chief of Staff
Global Q&A + read-only status queries with one-level dispatch to domain agents — one entry point for business users to command every capability.
Security and Trust by Design for Finance Scenarios
Light-confirmation safety model: agents autonomously reconnoiter, write, test and save; only execution requires one business-language confirmation
Every business side effect goes through the existing 'script + JobTask + BullMQ' pipeline — no second write path is invented
Tests never persist, script versions are rollbackable, and all tool calls are audited — financial data stays under control
SKILL.md skill standard + a five-category tool contract (read / write / test / trigger / validate) keeps capability boundaries explicit
Platforms & integrations
15 Major E-commerce Platforms — Domestic and Cross-Border Coverage
Domestic coverage spans JD POP, Douyin Shop, Pinduoduo, Xiaohongshu, WeChat Channels, VIP.com and the Alipay funds system; cross-border coverage spans Amazon, Temu, Shopee, Lazada, eBay, AliExpress, Shopify storefronts and Newegg. The platform integration framework is ready, with the full matrix rolling out progressively.
A built-in universal ERP integration framework auto-aligns supply-chain orders (three-code system) with platform statements; Kingdee Cosmic receivables join revenue reconciliation directly at line level — a closed loop across platform statements ↔ supply-chain orders ↔ ERP accounting.
About us
Guangdong Qingyiyun Software Technology Co., Ltd.
Our core product Qeasy focuses on enterprise data integration and application building. This intelligent reconciliation platform is Qeasy's flagship enterprise-grade AI Agent deployment: fact-data engineering as the foundation, a knowledge base as cognition, and an AI Agent matrix as execution — serving real financial reconciliation at large-scale e-commerce companies. Not a proof of concept, but a reconciliation pipeline running in production every month.
