Skip to content
Projects
Fullstack

Sales order approval AI agent

An AI-assisted review tool that cross-checks each pending sales order against the customer's purchase order, addresses, vendor pricing and margin before a person approves it.

Stack
  • React
  • TypeScript
  • Node.js
  • Express
  • OpenAI
  • PostgreSQL
  • MUI
Sales order approval AI agent

My role

As a fullstack developer in a team, I led the AI analysis and built most of the analysis interface, while teammates built the PDF report, the address analysis and the ERP approval step.

  • Analysis controller: Orchestrated document extraction and the parallel AI analyses for each order.
  • Line items analysis: Built the comparison of quantities, prices, totals and terms between documents and the system.
  • Vendor price history: Developed the AI-generated narrative of historical vendor pricing.
  • Lead-time tool: Implemented a logistics lead-time lookup exposed to the LLM through function calling.
  • Margin assessment: Built the deterministic margin health check in code.
  • Analysis UI: Developed most of the review screens in React and MUI.

Key project features

  • List of pending sales orders with the analysis status of each one.
  • Retrieval of the customer purchase order and vendor quote documents from the ERP, with LLM-based data extraction.
  • Three AI analyses running in parallel: line items consistency, billing/shipping addresses and vendor price history.
  • Strict JSON schema outputs for every AI response.
  • Persisted results with a generated PDF report stored in cloud storage, plus a batch job that pre-computes analyses.
  • Side-by-side document-vs-system comparison per item, with consistency status, accuracy and confidence.
  • Approval gated fail-closed by backend checks and always requiring a human click.

Impact achieved

The pre-approval review became faster and more consistent: reviewers see every discrepancy in one place instead of cross-checking documents by hand, while the final decision always stays with a person.

Technologies

Technology Use / Implementation
React Review interface with per-item comparison and approval flow.
TypeScript Typed code across frontend and backend.
Node.js Backend runtime for the analysis pipeline and batch jobs.
Express REST API for orders, analyses and approvals.
OpenAI Document extraction, AI analyses with structured outputs and function calls.
PostgreSQL Persistence of analyses and their results.
MUI Component library for the review screens.

Project flow

Sales order approval flow: document retrieval, LLM extraction, parallel AI analyses, margin check, review UI and human approval