Efekta R&D Blog // 24 September 2025
In 2023, we launched HyperClass, our virtual classroom platform designed to amplify what makes human teachers exceptional: clarity, adaptability, motivation, and meaningful feedback. Since then, we’ve delivered over 5.5 million lessons in this environment, gathering a uniquely rich dataset of teacher–student interaction.
“Can AI match — or enhance — the quality of human-led, synchronous teaching? ”
From Human-Centric Design to AI Instruction
HyperClass was designed to recreate (and in some ways improve upon) the physical classroom experience online. Its architecture blends the teacher’s webcam, interactive lesson tools, and immersive virtual environments, allowing live vision mixing and dynamic scene changes to guide attention and encourage participation.
With increasing demand for teachers and advances in large language models and orchestration systems, we saw an opportunity to adapt this environment for an AI instructor. Our goals were simple:
Function like a teacher — navigating lessons, using tools, and adapting to learner needs in real-time.
Behave like a teacher — establishing presence, building rapport, and responding to classroom dynamics.
Use the same environment — identical tools, content, and pedagogy to our human instructors.
Building the Teacher Model
We began by modelling the instructional behaviours of our expert teachers using our existing classroom AI toolset — a repository containing years of annotated teaching data, including lesson flow, tool usage, and student interaction patterns.
The system architecture involves:
A pedagogical model trained on human-teacher data to sequence activities, adjust explanations, and provide feedback.
An orchestration layer to manage time, navigate lesson scenes, and coordinate multiple AI agents.
Tool-use agents capable of manipulating HyperClass features — whiteboards, interactions, roleplay scenes — with the same flexibility as human teachers.
Multi-modal model observing how students use the free-form canvas and responds via text, voice, gestures, or whiteboard.
This approach preserves the context in which human teachers operate, allowing the AI to make tool-use decisions dynamically rather than following a fixed script.
Early Capabilities
In current tests, HyperTeacher can:
Navigate and manage entire lessons in HyperClass.
Switch tools and activities ad hoc in response to learner cues.
Simplify or expand explanations in real time.
Provide targeted, constructive feedback, remembering prior learner performance.
Maintain conversational presence — small talk, callbacks, and encouragement.
Notably, to better connect with a wide range of students, we intentionally avoided hyper-realistic avatars. Our research shows that overly realistic representations can reduce engagement (the “uncanny valley” effect) by leaving less room for the learner’s imagination to build connection.
Lessons Learned So Far
Context is critical. AI tool use is most effective when it mirrors human decision-making patterns, rather than following deterministic rules.
Pedagogy transfer works. Training the AI on the same instructional design principles we use for human teachers results in more coherent and effective lessons.
Multiple agents outperform monolithic models. Specialised sub-agents for orchestration, content delivery, and tool manipulation allow for more adaptive, human-like behaviour.
Next Steps
We are now expanding beyond single lessons to multi-lesson series enabling longitudinal learning and richer data feedback loops. Future work includes:
Extending activity types beyond the current HyperClass set.
Introducing new subjects into the platform, including STEM.
Integrating cultural and social aspects of learning.
Evaluating long-term learning outcomes compared to human-led classes.
HyperTeacher is still early in its development cycle, but initial results suggest that AI can inhabit a teaching role with surprising fidelity — especially when embedded in an environment purpose-built for human instruction. Our goal is not to replace teachers, but to make high-quality instruction available at scale, without compromising the pedagogical integrity of our teaching.
