Project detail
CleverAttendance OS: Autonomous Multi-Agent AI System
Engineered a production-ready, multi-agent AI customer support ecosystem during a 48h Microsoft Hackathon. It orchestrates a 6-agent LLM pipeline to autonomously resolve complex tickets, featuring a sentiment-driven "Human-in-the-Loop" fallback to eliminate user frustration.
AI Engineering • Semantic Kernel • Azure OpenAI • Human-in-the-Loop
The Problem: "Dumb Bots" and Frustrated Customers
Traditional AI chatbots fail at complex reasoning and lack empathy. When users hit a wall, they get trapped in endless loops, and when finally transferred to a human, the context is lost. The challenge was to build a system that acts intelligently but knows exactly when to step aside.
The Architecture: 6-Agent Orchestration Utilizing Microsoft Semantic Kernel and Azure OpenAI (GPT-4o mini), I architected a dynamic pipeline driven by a central Orchestrator agent.
Autonomous Resolution: The Orchestrator evaluates the case state and routes the flow through specialized agents (like ActionSupervisor and ActionCommander).
Deep Reasoning: The system executes up to 8 internal reasoning iterations before answering, ensuring complex issues are actually solved, not just deflected.
Robust Backend: An asynchronous Python polling loop integrates with a real-time PocketBase database to manage conversation states, all containerized into independent microservices via Docker.
The Magic: Emotional Intelligence & Multi-Tenant Scalability
Sentiment-Driven Human-in-the-Loop: The Receiver agent actively monitors emotional sentiment. If high stress or explicit human requests are detected, the HumanHandoff terminal agent takes control. It delivers an empathetic closing message and transfers a fully structured JSON context record to the human team. Nobody starts from zero.
Enterprise Scalability: Built as a Multi-Tenant SaaS. Business policies (like aggressive escalation rules) and API routes are dynamically injected via isolated JSON configurations, allowing multiple companies to run from a single deployment.
