Designing EDbot, a reflective companion for education at UM

What does it feel like to talk to EDbot?

When users start a conversation with EDbot, it doesn’t ask what tool they want to build or which problem they want to solve. Instead, it invites them into a different kind of dialogue, with questions such as:

• Who or what inspires you in education?
• What makes UM education meaningful and relevant for teachers, students, and society?
• What drives you as a student/teacher/researcher/staff?
• How will UM education look in 10 years?
• When is UM education at its best?

These questions are not random. They are meant to open up a reflective space that connects everyday teaching experiences to broader values, beliefs, and future aspirations.

EDbot listens carefully and responds with follow-up questions that help users dig a little deeper. The conversation moves gradually from concrete experiences to underlying motivations and ideas about purpose in education. There’s no rush, no right answer, and no expectation to perform, just time to think.

The AI component behind the scenes

The technical challenge in developing EDbot lay in orchestrating a multi-model architecture while keeping response times low and preserving the flow and purpose of the conversation. In traditional single-model chatbots, one generative AI model is often responsible for managing instructions, tracking user progress, and generating context-sensitive responses simultaneously. This "all-in-one" approach can lead to inconsistent user experiences and a a loss of the system's core intent.

To address this, I designed EDbot to distribute these responsibilities across independent, specialised models such as the Classification, Extraction, and Readiness models. While this separation provides greater control over the conversation, it also introduces background sub-processes that can increase the 'thinking time', or response latency. My task was to optimise this system so information could flow seamlessly between models in parallel. This ensured the system remained robust enough to capture deep, meaningful reflections while staying fast enough to feel natural and responsive for users. 

The system diagram below illustrates the sequential steps of the sub-processes in the technical architecture.

Step 1: the Classification model acts as a gatekeeper, instantly distinguishing between reflections on UM education and off-topic responses. It routes the dialogue either directly to the Extraction model to capture insights or to the Follow-up model to bring the conversation back on track.

Step 2: the Follow-up model acts as a gentle guide. Rather than correcting the user, it generates a polite bridge or a purposeful question that steers the dialogue back towards UM education, preserving the Socratic flow without breaking the momentum.

Step 3: the Extraction model scans user messages to identify and extract the core elements of their reflections, such as facts or opinions. By converting unstructured conversation into organised data points (topic, question, and answer), it builds a structured 'map' of the user's educational perspective and stores it in the Memory bank.

Step 4: the Memory bank acts as the central 'brain' of the dialogue. It ensures EDbot doesn't repeat a questions and forms the foundation for the final EduPersona, a visual summary of the conversation that reflects the user's educational values and vision.

Step 5: the Readiness model calculates a 'readiness score' after every turn. Based on a scoring rubric designed by the EDLAB team, this score determines whether the user has reflected on a sufficiently broad range of educational themes. 

Step 6: if the 'readiness score' is still low, the Chat model generates the next reflective question based on Korthagen’s layers of reflection. Drawing on the Memory bank, if focuses on probing "why?" rather than simply "what?".

Step 7: the Generation trigger is activated once the 'readiness score' is sufficiently high. At this point, users are invited to either continue the conversation or generate a personalised EduPersona.

Step 8: the Summarisation model acts as the storyteller, synthesising the full Memory bank into a cohesive first-person narrative and a set of visual instructions.

Step 9: the Image generation model uses these instructions to create a unique, pop-art style portrait. The result is the EduPersona, a visual representation of the user's educational vision and core values.

What does this mean for educators?

For educators, the technical design of EDbot exists primarily to support a focused and meaningful reflective conversation.

EDbot keeps the dialogue centred on education at UM, listens carefully to what users share, and monitors whether different aspects of their educational practice, such as inspiration, motivation, and future thinking, have been explored.

Why EDbot matters

For the EDLAB team and for myself as the technical builder, EDbot is an experiment in using AI differently. It isn't about optimising workflows or replacing human judgment, but about supporting reflection, meaning-making, and genuine conversation.

For me personally, this project was a journey in "human-centred engineering", learning how to design AI systems that ask better questions rather than provide faster answers. It showed how technical architecture can create a safe, reflective space where the value of AI lies not only in what it can do, but in what it helps us discover about ourselves.

By Dumitru Verșebeniuc

Contact me by email, if you have additional questions about EDbot.