Building an AI Voice TutorThat Actually Teaches
How we built Disco Party, a conversational AI that remembers what you learned yesterday and knows when you're ready to move on.
Disco Party
Disco Party is a voice AI tutor we built for students who want to learn on their own schedule. You talk to it like you're sitting with a real tutor - no typing, just conversation. It covers US history, government, social sciences, and other subjects based on American curricula.
This system provides a truly personalized learning experience, tracking your progress just like a real tutor. It incorporates material from previous lessons to continually assess and reinforce your understanding, with a visible progress bar that grows as you demonstrate mastery. If you get stuck or mix up a concept, the system walks you through it without just repeating the same explanation.
You can study for five minutes or spend an afternoon. When you close the app, it remembers where you were and what you understood. Next time you open it, you pick up exactly where you left off - Mr. Disco (your AI tutor) starts with a quick review based on your last session.
Real-Time Voice Tutoring
Real-time voice conversation with curriculum-driven questioning
Why Voice AI Tutoring Is Hard
Getting an AI to explain a concept isn't difficult. Claude, Deepseek and the GPT family can explain history or algebra just fine. The hard part is making it teach - not just answer questions on demand, but actually guide someone through learning over weeks or months.
Real tutoring means tracking what someone understands versus what they've just heard. It means maintaining context across sessions that might be days apart. It means knowing when to move forward and when to spend more time on fundamentals.
Most AI tutoring apps are basically ChatGPT with curriculum content bolted on. You ask questions, you get answers. But there's no real sense of progress, no memory of what you struggled with last Tuesday, no adaptation to how you actually learn.
We wanted to build something that worked more like an actual tutor.
Mastery-Based Learning
The system doesn't advance you because you clicked "next" or got one answer right. It moves you forward when your responses show you've actually understood the material well enough to build on it.
The system uses Agentic AI to manage the tutoring process. A tutor agent provides the personalized, real-time tutoring experience, utilizing multimodal models to deliver instructions. Simultaneously, an assessor agent continuously monitors the conversation, leveraging its memory of the student's past sessions to evaluate learning progress and provide targeted cues to the tutor agent. Both agents are supported by a Retrieval-Augmented Generation (RAG) system built on the student's textbooks, ensuring the content is accurate and curriculum-aligned.
Mastery is determined by the assessor agent looking for specific signals:
Consistency
Correctly answering similar or related concepts across multiple sessions.
Depth of Explanation
Moving beyond surface-level recall to provide comprehensive, nuanced explanations.
Concept Connection
Successfully linking the current topic to previously learned material without prompting.
The assessor agent uses these signals to inform the tutor agent when a student has demonstrated sufficient understanding to move on to the next concept or when they need to drill deeper into fundamentals before advancing.
If you go off topic - say you ask about World War II while working through a lesson on the American Revolution - Mr. Disco will either connect it back to what you're studying or offer to switch lessons if that's what you actually want to focus on.
Session continuity - the system remembers your progress and adapts
Admin Tools
The admin tools let non-technical users adjust how Mr. Disco teaches without rebuilding the system. Prompts can be updated, A/B tested, and tuned based on how students are actually learning. This matters because teaching isn't static - you need to iterate on approach without breaking the experience.
Model Routing & User Targeting
Admins choose which AI models handle live tutoring sessions and can route traffic between them. They can also target specific learner segments, like first-time users, highly active cohorts, or pilot classrooms, so prompt changes get tested on the right audience before a broader rollout.
Technical Infrastructure
The voice interaction runs on LiveKit, the same technology used in professional video conferencing and OpenAI's ChatGPT app. This handles the real-time audio processing without the delays or choppiness you get from cheaper solutions. When you speak, your words get processed immediately. When Mr. Disco responds, it sounds like an actual conversation, not a robot reading a script.
The whole system runs on Google Cloud, ensuring it works reliably whether you're studying at 3 PM or 3 AM, on your phone or desktop. We used serverless technology to achieve this result at low costs and without complex devops, ensuring fast shipping with modern practices. Furthermore, Google Cloud and Livekit dynamically route traffic to the nearest servers to your location to ensure global availability with low latency. Nothing heavy to download, no waiting for things to load; it just works seamlessly in your browser.
LiveKit
Real-time voice processing
Google Cloud
Serverless infrastructure
What We Learned
The most surprising and challenging technical hurdle was Prompt Engineering itself. Achieving the precise 'Mr. Disco' personality and educational experience required constant fine-tuning. Our team collaborated closely with the customer, not just to achieve the initial vision, but to empower them for the long term. We developed specialized tooling that allows non-technical users to A/B test prompt variations in a controlled environment. This capability was critical for rapid prototyping and fast iteration, ensuring the tutoring experience could evolve based on real-world student feedback without breaking the core system.
Conclusion
Disco Party shows what's possible when you build around how people actually learn instead of just adding a chat interface to existing content. Real-time voice conversation, persistent context across sessions, curriculum-driven progression, mastery-based advancement.
It is one project we've built at Jinka. We work with clients on AI applications across education, workflow automation, and other domains. If you're building something complex and need engineers who've solved similar problems, get in touch.
Building something complex?
We work with clients on AI applications across education, workflow automation, and beyond. Let's talk about what you're building.
Let's talk
