The latest articles on DEV Community by Nabil FATTOUCH (@nabilfattouch1). https://dev.to/nabilfattouch1 https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4018430%2F0a1878b1-4a83-453f-87a1-983c9e77c56e.jpg https://dev.to/nabilfattouch1 en Nabil FATTOUCH Mon, 06 Jul 2026 22:03:34 +0000 https://dev.to/nabilfattouch1/building-veqra-ai-how-i-resolved-enterprise-incidents-in-13-seconds-with-qwen3-235b-196d https://dev.to/nabilfattouch1/building-veqra-ai-how-i-resolved-enterprise-incidents-in-13-seconds-with-qwen3-235b-196d <p>VEQRA is an existing Microsoft 365 automation platform I built that detects and routes enterprise incidents. But it couldn't answer the critical questions: Why did this happen? What is the financial impact? What should we do right now?</p> <p>The Qwen Cloud Global AI Hackathon was the opportunity to build the intelligence layer VEQRA was missing.</p> <h2> What it does </h2> <p>VEQRA AI orchestrates three specialized AI agents that resolve a critical enterprise incident in 13 seconds:</p> <ul> <li>Memory Agent — searches historical incident database, identifies similar past cases, determines root cause with confidence score</li> <li>BI Agent — calculates financial impact, projects SLA breach risk, assigns criticality score</li> <li>Action Agent — generates structured action plan: Teams task, email to Data Owner, Power BI dashboard update</li> </ul> <p>Demo scenario: a critical Leasing VIP contract (€120,000 OVERDUE) is resolved in 13 seconds with zero human intervention.</p> <h2> How I built it </h2> <p>Each agent calls Qwen3-235B directly via Alibaba Cloud DashScope, using the OpenAI-compatible endpoint. Agents communicate through structured JSON outputs, coordinated sequentially by a Python orchestrator.</p> <h2> Challenges </h2> <p>Designing prompts that produce consistent, structured JSON outputs across all three agents was the hardest part — keeping the total resolution time under 15 seconds required careful prompt engineering.</p> <h2> What I learned </h2> <p>Depth beats breadth. One perfect scenario executed flawlessly is more valuable than five agents that partially work. Qwen3-235B's long-context reasoning and structured output capabilities made the multi-agent orchestration reliable and deterministic.</p> <p>GitHub: <a href="https://github.com/nabilfattouch1/VEQRA-AI" rel="noopener noreferrer">https://github.com/nabilfattouch1/VEQRA-AI</a><br> Demo: <a href="https://youtu.be/zakd6bsdzDA" rel="noopener noreferrer">https://youtu.be/zakd6bsdzDA</a></p> ai python opensource machinelearning