AIdeas: Tutor AI, a LLM powered adaptive learning platform to help students catch up

Education systems around the world are designed around standardized progression. Curricula assume that students move through subjects at a uniform pace and with similar foundations. In reality, learning rarely works this way. Students often move forward with partial understanding, missing prerequisites, or uneven exposure to key concepts. Students who have less access to education, often fall behind at the foundation level and are discouraged from following their passion. I have seen this happen to several students of different age groups during my time as a volunteer tutor in India, US, and Canada.

My vision is to use artificial intelligence to create a personal tutor that identifies and addresses these gaps in knowledge. The goal is not to replace existing education systems, but to complement them by providing structured guidance that adapts to each learner’s starting point.

The system assesses a student’s current level, identifies missing prerequisites, and builds a learning pathway tailored to their goals. Instead of producing entirely new educational content, the platform curates and recommends high quality, pre indexed learning materials, beginning with sources such as YouTube. It then evaluates comprehension through short assessments and adjusts the pathway accordingly.

Over time, the system learns which learning resources work best for different types of students. This creates a feedback loop where both the learning pathways and the resource recommendations improve continuously.

The broader ambition is to make structured, personalised learning accessible to anyone with an internet connection.
__________________________________________________________________________________

Across many education systems, students fall behind not because of a lack of ability, but because of missing foundations. Once a prerequisite concept is unclear, subsequent topics become progressively harder to understand. Without intervention, these gaps compound over time.

This problem is particularly visible in technical fields such as programming, data science, and machine learning. A student may be interested in studying machine learning but may never have had the opportunity to learn Python or statistics in school. When they attempt to enter the field directly, the gap between expectation and preparation can be discouraging.

The challenge is even more pronounced for students in smaller towns, under resourced schools, or communities where specialized subjects are not widely available. Many motivated learners attempt to self study using freely available online resources, but the sheer volume of material often leads to confusion rather than clarity. Students struggle to determine where to start, what to learn next, and how to evaluate their progress.

Language can also be a barrier. A large portion of high quality educational content is created in English, which may not be the primary language for many learners globally. Bilingual or non native English speakers often require resources presented in ways that better match their learning context.

An adaptive AI tutor can address several of these challenges. By diagnosing a learner’s current level, sequencing the right concepts, and recommending resources that match both their knowledge and language context, the system can reduce the friction that many learners experience when attempting to study independently.

The result is not simply improved efficiency in learning. It is increased access to fields that would otherwise feel inaccessible.

The system was developed using a structured engineering approach that separated the platform into modular components.

To accelerate development, I used Kiro Code to scaffold the project and organize the implementation workflow. The process began with the creation of steering documents that defined the system architecture, learning workflow, and data structures required for student tracking. From these documents, I generated detailed task lists that guided the implementation process.

The first component I built was the content indexing pipeline.

Educational videos were ingested and processed through a tagging system that identifies topics, subtopics, and prerequisite relationships. Each video and its segments were converted into embeddings and stored in a vector database. This allows the platform to perform semantic retrieval when recommending learning material.

The indexing system effectively converts a large set of educational videos into a searchable knowledge base organized by concept rather than by playlist or channel.

The next step was implementing the student management system.

Each learner has a persistent profile that tracks their goals, historical performance, and conceptual understanding. The platform also stores longer form session notes that document progress over time. These records are maintained as structured markdown files that provide a chronological learning history for each student.

Once the student layer was established, I implemented the quiz generation system. After a student consumes recommended learning material, the platform generates short quizzes to evaluate comprehension of the specific concept being studied. The results provide a signal about whether the learner has achieved sufficient mastery.

With these pieces in place, I built the suggestion engine that orchestrates the learning flow.

The engine takes three inputs:

Using these inputs, the system selects the most relevant next learning resource from the pre indexed video library. If the student demonstrates mastery, the engine progresses to the next concept. If the student struggles, the system recommends alternative explanations or supporting material.

After each recommendation, the platform generates a follow up quiz and repeats the evaluation cycle.

In later iterations, I expanded the suggestion engine to track patterns in learning outcomes. When a student repeatedly struggles with a particular type of resource, the system updates the learner profile to capture potential learning preferences. This allows the recommendation process to adapt over time.

I also began experimenting with bilingual capabilities. The goal is to recommend educational resources in multiple languages when available. While the indexing framework supports this capability, building a reliable multilingual content library requires further work.

__________________________________________________________________________________________
Demo: https://www.youtube.com/watch?v=NOpp2ELh4KE

The platform is built using a cloud-native architecture on AWS designed to support scalable adaptive learning.

Students interact with the platform through a web application delivered via Amazon CloudFront and hosted on Amazon S3. Backend application services run on container infrastructure using Amazon ECS or Elastic Beanstalk, with authentication managed through Amazon Cognito.

The core learning workflows are implemented through serverless services. AWS Lambda functions handle key tasks such as educational content indexing, quiz generation, recommendation logic, and student profile updates. These functions are orchestrated using AWS Step Functions to manage the learning pipeline and asynchronous processing.

Artificial intelligence capabilities are powered through Amazon Bedrock. Bedrock models are used to generate quizzes, analyze student performance, and support multilingual learning scenarios. Embedding models are also used to generate semantic representations of indexed learning content.

The data layer combines structured and unstructured storage systems. Amazon RDS PostgreSQL stores student profiles, course goals, and quiz results. A vector database built on pgvector or OpenSearch enables semantic retrieval of educational videos based on conceptual similarity. Amazon S3 stores larger learning artifacts such as student progress logs, session notes, and indexed educational transcripts.

Together, these components form an adaptive learning feedback loop in which the platform continuously evaluates student understanding and dynamically recommends the next learning resource.

Learnings

Developing this platform provided several insights into both modern AI development workflows and the practical design of adaptive learning systems.

One of the most valuable aspects of the development process was the combination of Kiro for AI-assisted software development and the breadth of services available within AWS. Kiro provided a structured way to translate high-level system ideas into concrete engineering tasks. By generating steering documents and implementation plans, it helped establish an organized development workflow from the outset.

At the infrastructure level, AWS provided a natural environment for building a modular and scalable architecture. Services such as AWS Lambda enabled lightweight serverless execution for indexing pipelines and recommendation logic, while Amazon RDS and S3 provided clear separation between structured data and larger learning artifacts. Amazon Bedrock also simplified the integration of foundation models by providing managed access to large language models and embedding capabilities without the operational overhead of hosting models directly.

Together, Kiro and AWS created a development environment that allowed rapid experimentation while maintaining architectural discipline.

Another important lesson came from working with spec driven development within Kiro. The approach relies on clearly defined steering documents that describe system components, interfaces, and workflows before code is written.

This approach proved particularly effective when tasks were well scoped and relatively contained. When the problem space was narrow, Kiro could translate specifications into useful code structures and implementation tasks with a high degree of reliability.

However, when specifications became overly complex or attempted to define too many components simultaneously, the quality of generated outputs declined. Excessive steering sometimes produced implementations that were technically correct but poorly aligned with the intended architecture.

This led to an important development practice: keeping specifications concise and decomposing large problems into smaller units of work. Smaller tasks produced more consistent results and were easier to validate.

Another observation was that different models perform better on different categories of tasks. Certain models were particularly effective at structured engineering work, while others were more suitable for simpler transformations or content generation. Selecting the right model for each stage of development became an important part of the workflow.

The project also provided practical experience in prompt design and model behavior.

Different language models respond differently to instructions, even when the underlying task is similar. Some models perform best with highly structured prompts that define expected output formats and constraints. Others respond better to more flexible instructions that allow them to reason through the problem.

Learning how to adjust prompts for different models became an important part of achieving consistent results. This included specifying the format of outputs, defining evaluation criteria, and carefully controlling the scope of each request.

Over time, this led to a better understanding of how prompting can function as a form of interface design between human developers and AI systems.

Finally, building the recommendation and suggestion engine provided insight into the practical challenges of adaptive learning systems.

The effectiveness of a tutoring system depends not only on the quality of the educational content but also on how learning progress is evaluated. Short quizzes provided a reliable signal for determining whether a student had understood a concept, allowing the system to adjust recommendations dynamically.

Another important factor was maintaining a persistent learning history for each student. Tracking what resources were recommended, how students performed on follow up quizzes, and which materials proved ineffective allowed the platform to refine future recommendations.

The process highlighted how recommendation systems in educational contexts differ from typical content recommendation engines. Instead of optimizing for engagement, the objective is to guide learners through prerequisite structures while minimizing frustration and knowledge gaps.

Understanding these design considerations will be essential as the system expands to support additional subjects, languages, and types of learning resources.