Describe your project

The project aims to develop an intelligent tutoring system that leverages a Large Language Model (LLM) to function as a full-fledged adaptive teacher. The system will provide tailored instruction to students by dynamically adjusting the difficulty and depth of explanations based on their level of understanding, personal learning preferences, and psychological needs. The solution will focus on keeping the student engaged, resolving doubts in real-time, and supporting them throughout their learning journey.

1. In-Scope of the Solution:
   Master Prompt and Dynamic Interaction:

The system will use an LLM to generate explanations for a specific subject or chapter. A master prompt will define the overall lesson plan, but the system will adjust its response based on student interactions.
When a student asks questions or expresses confusion, the LLM will not only respond to the immediate doubt but also adjust the explanation level (simpler or more detailed) to match the student's current understanding.
Adaptive Questioning and Content Delivery:

The system will assess the student's comprehension level in real-time, providing questions such as multiple-choice (MCQs), long-form questions, and conceptual queries to reinforce learning.
If the student struggles, the system will reframe explanations in simpler terms, using analogies or conversational examples tailored to the student's understanding.
Personalization and Child Psychology:

The system will account for child psychology and different learning styles, tuning explanations based on student feedback (verbal or non-verbal). It will focus on maintaining engagement and encouraging active learning by personalizing the teaching style according to the child's needs.
Continuous monitoring and modification of the master prompt to reflect evolving student progress.

Challenges we ran into

1. Dynamic Response Personalization
   Issue: One of the biggest challenges we faced was ensuring that the system could dynamically adjust its responses based on the student’s input. The LLM would sometimes generate responses that were too complex or too simple for the student’s level of understanding.
   Solution: We implemented a feedback loop where the system evaluates each student response to detect comprehension levels. We also defined several "response tiers" (beginner, intermediate, advanced) and used heuristics or explicit user feedback to guide the system in selecting the appropriate tier for each response.
2. Maintaining Context Over Long Conversations
   Issue: During long tutoring sessions, the LLM sometimes lost track of context, leading to irrelevant or repetitive explanations. This made it harder for the system to provide continuity in learning.
   Solution: We employed a session management system that stored key information from each interaction, like the student’s current knowledge level and past questions. This context was regularly fed back into the LLM to ensure continuity. We also fine-tuned prompts to maintain focus on the current topic and prevent drifting off-topic.
3. Detecting When a Student Is Confused
   Issue: It was difficult for the system to accurately gauge when a student was confused or struggling. While incorrect answers were a clear sign, sometimes the student might have been hesitant or unclear in their questions, and the system would fail to simplify the explanation adequately.
   Solution: We incorporated more explicit conversational cues from the LLM, where it regularly asked the student how comfortable they felt with the explanation ("Did that make sense?" or "Would you like me to explain that in another way?"). Additionally, we created a system to track repeated queries or incorrect responses as potential indicators of confusion, prompting the system to simplify or rephrase explanations automatically.