VCE English Language Online

Main Takeaways

• Implemented OpenAI's function calling system to create an AI assistant that can modify its own UI and behavior
• Designed a dynamic personality system that enhances learning by demonstrating different linguistic registers and styles
• Created a stateful middleware layer that bridges client-side and server-side actions for seamless user experience

Introduction and Background

As a high school teacher with a degree in linguistics from Melbourne University, I recognized the need for a resource to support students and teachers in understanding the metalanguage of the VCE English Language course. I created VCE English Language Online to address this need, leveraging my expertise in linguistics to develop a comprehensive online platform. The technical architecture was built on Next.js with React for the frontend and custom API routes handling the AI integration. I chose this stack for its server-side rendering capabilities, and integration of front and back-end components

AI-Based English Language Essay Generator

I included a React application that utilized OpenAI's API to generate essays in the format and style expected by the course examiners. The application allowed students to input modern language quotes and pieces of media information, which were then passed through an AI API route to generate an essay. A second API call was made to assess the generated essay against the exam assessment standards and linguistics content, providing feedback to students. To ensure reliable essay generation, I implemented a prompt engineering system with specific constraints and examples of high-scoring essays. The architecture included error handling with fallback content for API failures and a caching layer to reduce token usage for similar prompts, optimizing both performance and cost efficiency while safeguarding against potential issues with AI hallucinations.

AI Chat Assistant and Metalanguage Support

To support students in understanding the specific metalanguage of the course, I created an AI chat assistant trained on the course material and metalanguage. The implementation utilized OpenAI's GPT-4o-mini model wrapped in a custom Next.js API route that handled message formatting, prompt injection, and function calling. I structured the system using a two-stage API call approach: an initial call to determine if any functions needed to be executed, followed by a second call incorporating function results. This architecture gave the assistant the ability to access metalanguage definitions through the `getAllMetalanguage` and `getCourseTermDefinition` functions, as seen in the API route code. I also incorporated engaging features, such as an 8-bit style design and hidden Easter eggs, to make the chatbot an enjoyable and fun experience for students.

Personalization and Register Support

I utilized OpenAI's function calling capabilities to enable the chatbot to regenerate its UI and change its personality in response to user requests. As shown in the provided API route code, the chatbot could adopt different personality styles through the `changePersonality` function, each with its own unique register and speech style. The system implemented a personality state that persisted across messages, with the `getPersonalityPrompt` function dynamically adjusting the system prompt for each conversation turn. This feature allowed students to explore different registers and dialects, gaining a deeper understanding of the course material. Technical implementation included a state management system that tracked the current personality across requests and a DOM manipulation function (`changeChatColor`) that adjusted the UI based on which personality was active.

Technical Implementation and Challenges

The development of the platform presented several technical challenges, particularly in the coordination between client and server operations. My API route code shows how I addressed this by implementing a sophisticated function calling system where messages were first processed to determine required actions, then functions were executed server-side, and finally results were incorporated into a second API call. This approach allowed the AI to trigger UI changes (e.g. through `changeChatColor`), access course-specific data (via `getAllMetalanguage`), and modify its own behavior (with `changePersonality`). I structured the API route with proper error handling that provided detailed logging for troubleshooting while returning user-friendly error messages. The implementation also included content filtering with regex replacement (`assistantReply.replace(/<link>(.*?)</link>/g, '$1')`) to ensure proper formatting of responses before they reached the client.

Outcome and Impact

The VCE English Language Online platform was well-received by students and teachers, with many reporting that it had made it easier to understand the course material. The platform's popularity also led to tutoring clients and positive feedback from students and teachers, demonstrating the effectiveness of the AI-powered learning platform in supporting student learning and understanding of the VCE English Language course. The technical implementation proved robust, handling thousands of interactions with minimal downtime and consistent response quality. Analytics showed that students particularly engaged with the personality-switching feature, with the 'Gen Z' and 'Aussie' personas being most popular, indicating that the technical complexity of implementing this feature was justified by its educational value in demonstrating linguistic concepts through practical examples.