💡 Inspiration
Customer support and sales teams spend hundreds of hours manually reviewing unstructured customer emails - handling quote requests, negotiating volume discounts, answering order status queries, and processing downtime refund claims. Traditional rule-based chatbots fail because customer emails are free-form, filled with soft requests (e.g., "the quote looks high, can we get a more favorable rate given our volume?"), and can get hairy. On the other hand, fully autonomous AI bots risk over-promising discounts or approving unauthorized refunds without supervision. We built AutoOps to solve this: an end-to-end Autopilot AI Agent with Human-in-the-Loop Safeguards, powered by Qwen-2.5 on Alibaba Cloud. It parses unformatted customer emails, retrieves live context, executes business actions, and safely halts for human approval whenever commercial safety limits are reached.
──────
⚙️ What it does
• Intelligent Free-Form Email Ingestion: Monitors incoming and outgoing customer emails via IMAP/SMTP without enforcing any rigid structure or template.
• Automatic Database Schema Parsing: Connects to the database and automatically parses the schema for understanding available inventory and order data.
• Unified Agentic Reasoning & RAG: Uses Qwen-2.5 to evaluate customer emails alongside prior conversation thread history, live product catalogs, and PostgreSQL database order records.
• Implicit & Negotiation Parsing: Detects soft discount requests ("favorable rate", "volume consideration") and calculates optimal pricing proposals.
• Safe-Limits Settings: Allows users to set safe limits for transactions, and flag risky ones.
• Autopilot vs. Human-in-the-Loop (HITL) Safeguards: Evaluates every transaction against strict commercial policy caps.
• Safe Transactions (e.g., standard quotes, low-value refunds, stock-checked orders): Auto-approved and dispatched instantly.
• High-Risk Transactions (e.g., discounts >15%, refunds >50,stockshortages): Automatically routed to the interactive dashboard for 1-click sales manager review and customization.
• Alibaba Cloud OSS Audit Trail: Every diagnostic trace log and structured invoice record can be archived to Alibaba Cloud Object Storage Service (OSS).
──────
🛠️ How we built it
• AI Core: Qwen-2.5 (qwen-plus) accessed via Alibaba Cloud Model Studio (DashScope API) for intent classification, parameter extraction, and email response synthesis.
• Cloud Infrastructure: Deployed on Alibaba Cloud ECS (Elastic Compute Service) using Docker & Docker Compose.
• Cloud Storage: Alibaba Cloud OSS for audit log persistence and invoice archiving.
• Backend Services: Node.js & Express API engine with PostgreSQL (pg) for dynamic Text-to-SQL database queries, and Nodemailer / ImapFlow for real-time mailbox integration.
• Frontend Dashboard: Sleek, dark-mode web application (HTML5, CSS3, Vanilla JS) showing real-time execution logs, HITL approval queues, and outbox simulations.
──────
🚩 Challenges we ran into
Handling Unstructured Negotiation Language: Extracting business intent from soft human phrasing like "we were hoping for a more favorable rate" without explicit percentage metrics. We solved this by designing a generalized RAG prompt that passes prior thread history and catalog limits into Qwen-2.5.
State & Thread Continuity: Preventing the agent from treating follow-up negotiation emails as new, isolated quote requests. We implemented customer history RAG that injects prior quotes into the agent's reasoning loop.
──────
🏆 Accomplishments that we're proud of
• Customizable and Flexible Pipeline: Replaced rigid logic with a single, generalized Agentic Processing Pipeline that handles quote requests, renegotiations, order tracking, and refund disputes out of the box.
• HITL Safety Engine: Successfully balanced high automation speed with enterprise safety limits, ensuring AI never over-grants discounts or unauthorized payouts.
• Full Alibaba Cloud Stack Integration: Built a production-ready cloud deployment combining Qwen-2.5 (DashScope), Alibaba Cloud OSS, and containerized Alibaba Cloud ECS hosting.
──────
📚 What we learned
• Effective architectural patterns for designing safe, human-in-the-loop autonomous agents in mission-critical business applications.
──────
🔮 What's next for AutoOps
• Multi-Channel Integration: Expanding beyond email to support Slack, WhatsApp Business, ZenDesk, and ServiceNow tickets.
• ERP & CRM Sync: Deep integration with Salesforce and SAP for automatic contract generation and inventory reservation.
Log in or sign up for Devpost to join the conversation.