Ani Baghdasaryan

This talk focuses on how TUMO’s AI Colearner platform was engineered, covering architectural strategies, techniques for shaping LLM behavior, and the integration of external tools within the learning loop, as well as the mechanisms that enable real-time adaptation of interaction and feedback. We will explore how the system interprets student interaction signals-such as engagement, hesitation, or incorrect reasoning-to decide when to prompt learners to articulate their understanding versus when to intervene with structured, step-by-step guidance. These decisions are implemented through a modular LLM-based architecture that separates concerns such as dialogue control, task progression, and tool usage, allowing for more predictable and controllable behavior.
We will also discuss key engineering challenges encountered in production, including managing prompt sensitivity, maintaining consistent behavior across diverse scenarios, preventing over-scaffolding that can hinder learning, and enforcing usage constraints without degrading user experience. Particular attention will be given to how we achieved stable, reliable interactions at scale despite the inherent variability of large language models, and the tradeoffs involved in balancing flexibility, control, and pedagogical effectiveness.