Create your own custom extension for Azure Bicep
October 20, 2025Azure Managed Redis Deep Dive
October 21, 2025
Personal Background
My name is Peace Silly. I studied French and Spanish at the University of Oxford, where I developed a strong interest in how language is structured and interpreted. That curiosity about syntax and meaning eventually led me to computer science, which I came to see as another language built on logic and structure. In the academic year 2024–2025, I completed the MSc Computer Science at University College London, where I developed this project as part of my Master’s thesis.
Project Introduction
Can large-scale event management be handled through a simple chat interface?
This was the question that guided my Master’s thesis project at UCL. As part of the Industry Exchange Network (IXN) and in collaboration with Microsoft, I set out to explore how conversational interfaces and autonomous AI agents could simplify one of the most underestimated coordination challenges in campus life: managing events across multiple departments, societies, and facilities.
At large universities, event management is rarely straightforward. Rooms are shared between academic timetables, student societies, and one-off events. A single lecture theatre might host a departmental seminar in the morning, a society meeting in the afternoon, and a careers talk in the evening, each relying on different systems, staff, and communication chains. Double bookings, last-minute cancellations, and maintenance issues are common, and coordinating changes often means long email threads, manual spreadsheets, and frustrated users.
These inefficiencies do more than waste time; they directly affect how a campus functions day to day. When venues are unavailable or notifications fail to reach the right people, even small scheduling errors can ripple across entire departments. A smarter, more adaptive approach was needed, one that could manage complex workflows autonomously while remaining intuitive and human for end users.
The result was the Event Management Multi-Agent System, a cloud-based platform where staff and students can query events, book rooms, and reschedule activities simply by chatting. Behind the scenes, a network of Azure-powered AI agents collaborates to handle scheduling, communication, and maintenance in real time, working together to keep the campus running smoothly.
The user scenario shown in the figure below exemplifies the vision that guided the development of this multi-agent system.
Starting with Microsoft Learning Resources
I began my journey with Microsoft’s tutorial Build Your First Agent with Azure AI Foundry which introduced the fundamentals of the Azure AI Agent Service and provided an ideal foundation for experimentation. Within a few weeks, using the Azure Foundry environment, I extended those foundations into a fully functional multi-agent system.
Azure Foundry’s visual interface was an invaluable learning space. It allowed me to deploy, test, and adjust model parameters such as temperature, system prompts, and function calling while observing how each change influenced the agents’ reasoning and collaboration. Through these experiments, I developed a strong conceptual understanding of orchestration and coordination before moving to the command line for more complex development later.
When development issues inevitably arose, I relied on the Discord support community and the GitHub forum for troubleshooting. These communities were instrumental in addressing configuration issues and providing practical examples, ensuring that each agent performed reliably within the shared-thread framework.
This early engagement with Microsoft’s learning materials not only accelerated my technical progress but also shaped how I approached experimentation, debugging, and iteration. It transformed a steep learning curve into a structured, hands-on process that mirrored professional software development practice.
A Decentralised Team of AI Agents
The system’s intelligence is distributed across three specialised agents, powered by OpenAI’s GPT-4.1 models through Azure OpenAI Service. They each perform a distinct role within the event management workflow:
- Scheduling Agent – interprets natural language requests, checks room availability, and allocates suitable venues.
- Communications Agent – notifies stakeholders when events are booked, modified, or cancelled.
- Maintenance Agent – monitors room readiness, posts fault reports when venues become unavailable, and triggers rescheduling when needed.
Each agent operates independently but communicates through a shared thread, a transparent message log that serves as the coordination backbone. This thread acts as a persistent state space where agents post updates, react to changes, and maintain a record of every decision.
For example, when a maintenance fault is detected, the Maintenance Agent logs the issue, the Scheduling Agent identifies an alternative venue, and the Communications Agent automatically notifies attendees. These interactions happen autonomously, with each agent responding to the evolving context recorded in the shared thread.
Interfaces and Backend
The system was designed with both developer-focused and user-facing interfaces, supporting rapid iteration and intuitive interaction.
The Terminal Interface
Initially, the agents were deployed and tested through a terminal interface, which provided a controlled environment for debugging and verifying logic step by step. This setup allowed quick testing of individual agents and observation of their interactions within the shared thread.
The Chat Interface
As the project evolved, I introduced a lightweight chat interface to make the system accessible to staff and students. This interface allows users to book rooms, query events, and reschedule activities using plain language.
Recognising that some users might still want to see what happens behind the scenes, I added an optional toggle that reveals the intermediate steps of agent reasoning. This transparency feature proved valuable for debugging and for more technical users who wanted to understand how the agents collaborated.
When a user interacts with the chat interface, they are effectively communicating with the Scheduling Agent, which acts as the primary entry point. The Scheduling Agent interprets natural-language commands such as “Book the Engineering Auditorium for Friday at 2 PM” or “Reschedule the robotics demo to another room.” It then coordinates with the Maintenance and Communications Agents to complete the process.
Behind the scenes, the chat interface connects to a FastAPI backend responsible for core logic and data access. A Flask + HTMX layer handles lightweight rendering and interactivity, while the Azure AI Agent Service manages orchestration and shared-thread coordination. This combination enables seamless agent communication and reliable task execution without exposing any of the underlying complexity to the end user.
Automated Notifications and Fault Detection
Once an event is scheduled, the Scheduling Agent posts the confirmation to the shared thread. The Communications Agent, which subscribes to thread updates, automatically sends notifications to all relevant stakeholders by email. This ensures that every participant stays informed without any manual follow-up.
The Maintenance Agent runs routine availability checks. If a fault is detected, it logs the issue to the shared thread, prompting the Scheduling Agent to find an alternative room. The Communications Agent then notifies attendees of the change, ensuring minimal disruption to ongoing events.
Testing and Evaluation
The system underwent several layers of testing to validate both functional and non-functional requirements.
Unit and Integration Tests
Backend reliability was evaluated through unit and integration tests to ensure that room allocation, conflict detection, and database operations behaved as intended. Automated test scripts verified end-to-end workflows for event creation, modification, and cancellation across all agents. Integration results confirmed that the shared-thread orchestration functioned correctly, with all test cases passing consistently. However, coverage analysis revealed that approximately 60% of the codebase was tested, leaving some areas such as Azure service integration and error-handling paths outside automated validation. These trade-offs were deliberate, balancing test depth with project scope and the constraints of mocking live dependencies.
Azure AI Evaluation
While functional testing confirmed correctness, it did not capture the agents’ reasoning or language quality. To assess this, I used Azure AI Evaluation, which measures conversational performance across metrics such as relevance, coherence, fluency, and groundedness.
The results showed high scores in relevance (4.33) and groundedness (4.67), confirming the agents’ ability to generate accurate and context-aware responses. However, slightly lower fluency scores and weaker performance in multi-turn tasks revealed a retrieval–execution gap typical in task-oriented dialogue systems.
Limitations and Insights
The evaluation also surfaced several key limitations:
- Synthetic data: All tests were conducted with simulated datasets rather than live campus systems, limiting generalisability.
- Scalability: A non-functional requirement in the form of horizontal scalability was not tested. The architecture supports scaling conceptually but requires validation under heavier load.
Despite these constraints, the testing process confirmed that the system was both technically reliable and linguistically robust, capable of autonomous coordination under normal conditions. The results provided a realistic picture of what worked well and what future iterations should focus on improving.
Impact and Future Work
This project demonstrates how conversational AI and multi-agent orchestration can streamline real operational processes. By combining Azure AI Agent Services with modular design principles, the system automates scheduling, communication, and maintenance while keeping the user experience simple and intuitive.
The architecture also establishes a foundation for future extensions:
- Predictive maintenance to anticipate venue faults before they occur.
- Microsoft Teams integration for seamless in-chat scheduling.
- Scalability testing and real-user trials to validate performance at institutional scale.
Beyond its technical results, the project underscores the potential of multi-agent systems in real-world coordination tasks. It illustrates how modularity, transparency, and intelligent orchestration can make everyday workflows more efficient and human-centred.
Acknowledgements
What began with a simple Microsoft tutorial evolved into a working prototype that reimagines how campuses could manage their daily operations through conversation and collaboration. This was both a challenging and rewarding journey, and I am deeply grateful to Professor Graham Roberts (UCL) and Professor Lee Stott (Microsoft) for their guidance, feedback, and support throughout the project.