Skip to content
Projects
Backend

Walle logistics AI agent

An AI agent that reads vendor email threads about open purchase orders, summarizes and classifies them, and turns them into prioritized, auditable actions for logistics analysts.

Stack
  • TypeScript
  • Node.js
  • Express
  • OpenAI
  • PostgreSQL
  • Azure Web PubSub
  • Power Automate
Walle logistics AI agent

My role

As a co-developer in a team, I focused on turning the agent's analysis into real actions. The AI classification and suggestion core was built by teammates; my work covered the execution layer and several modules around it.

  • Action execution layer: Built the execution of vendor emails, Teams notifications and ERP updates.
  • Authentication and roles: Implemented user authentication and role-based access.
  • Standard actions and prioritized lists: Developed the deterministic standard actions per scenario and the prioritized daily worklists.
  • Multilingual notifications: Built the notification messages in multiple languages.
  • Finance module: Developed the module for payment supports.
  • International transit: Built most of the international transit workflow: pickups, freight quotes, customs follow-up and last mile.

Key project features

  • Daily orchestrated sync that pulls pending vendor emails from the ERP and cleans them.
  • LLM-generated structured summary and classification of the latest message into predefined scenarios, with confidence.
  • Standard actions per scenario plus AI-suggested actions (vendor email, Teams notification, PO status/date updates) with bilingual messages and quoted evidence.
  • Code guardrails that discard actions for auto-replies and irrelevant mail.
  • SQL views that prioritize a daily worklist; an analyst reviews and executes each action.
  • Full audit of suggested versus manual actions, real-time notifications and per-call token/cost tracking.

Impact achieved

Analysts no longer have to read long vendor email threads from scratch every day: they start from a prioritized worklist with summaries and suggested actions, execute them in one place and keep every decision audited.

Technologies

Technology Use / Implementation
TypeScript Typed code across the backend.
Node.js Runtime for the orchestrator and the action execution layer.
Express REST API for worklists, actions and modules.
OpenAI Email summaries, scenario classification and suggested actions.
PostgreSQL Storage, prioritization views and audit log.
Azure Web PubSub Real-time notifications to the analysts.
Power Automate Workflow automation integrated with the agent.

Project flow

Walle flow: ERP vendor emails, daily orchestrator, LLM summary and classification, standard and suggested actions, guardrails, analyst review and action execution