Agent-Based Scheduling System
Project Snapshot
- Category: Generative AI / Agentic Systems
- Focus: Agent orchestration, workflow control, scheduling automation
- Architecture: Event-driven, ReAct-based, tool-augmented agents
- GitHub: Agent-Based Scheduling System
Executive Summary
Agent-Based Scheduling System is a production-oriented, agentic workflow designed to automate doctor discovery and appointment coordination using structured reasoning and controlled tool execution. Unlike traditional rule-based scheduling pipelines, the system dynamically interprets user intent, selects appropriate actions, and produces traceable side effects through an event-driven agent architecture.
Problem Statement
Scheduling workflows in operational systems are often rigid and brittle, leading to:
- Inability to handle ambiguous or multi-intent user requests
- Hard-coded decision logic that does not generalize well
- Lack of transparency into automated decisions
- Poor observability and limited evaluation of system behavior
Solution Overview
I designed this system as an agent-based, event-driven workflow that separates reasoning from action, enabling flexible decision-making while maintaining full control over side effects.
- LLM-based intent detection and decision-making using the ReAct pattern
- Semantic doctor search via vector embeddings and retrieval
- Explicit tool invocation for scheduling actions
- Deterministic execution with safe termination guarantees
Architecture & Approach
- Implemented using LlamaIndex event-driven workflows
- Explicit workflow events (prep, input, tool call, stop)
- Clear separation between LLM reasoning and state-changing actions
- Structured agent traces for observability and debugging
- Evaluation hooks integrated directly into the agent lifecycle
Key Capabilities
• Event-Driven Agent Workflow
Uses explicit workflow events to control execution flow, prevent runaway loops, and ensure deterministic termination.
• Semantic Doctor Discovery
Performs embedding-based semantic search over structured doctor profiles instead of relying on keyword matching.
• Controlled Tool Execution
All state changes (scheduling requests) are executed through explicitly defined tools, ensuring safety and auditability.
• Agent Traces & Evaluation
Logs intent detection, tool selection, execution outcomes, and latency, enabling behavior-based evaluation and monitoring.
Impact & Outcomes
- Demonstrated reliable agentic scheduling without hard-coded flows
- Improved transparency through structured decision traces
- Enabled evaluation of agent behavior beyond text-based metrics
- Created a reusable blueprint for workflow-driven AI automation
Tech Stack
Languages: Python
Agent Framework: LlamaIndex (Workflows, ReAct)
LLMs: Groq-hosted LLaMA models
Embeddings: HuggingFace (BGE family)
API: FastAPI
Deployment: Docker
Evaluation: Intent accuracy, tool selection accuracy, task completion, latency