Mr.So

Inspiration

In the era of Generative AI, we are often overwhelmed by "answers" but starved for "understanding." Traditional LLM interactions frequently lead to passive reading where information is quickly forgotten. We were inspired by the Socratic Method, a form of cooperative argumentative dialogue that stimulates critical thinking. We wanted to build a mentor that doesn't just give you the fish but teaches you how to fish by asking the right questions.

What it does

Mr. So is an AI-native iOS application that transforms passive consumption into active mastery. By leveraging Gemini 1.5 Pro, it turns any topic into a structured, interactive learning journey:

Structured Curriculum: Generates a 3-stage learning path tailored to the user's proficiency level (Beginner to Advanced).

Socratic Dialogue: Powered by Gemini’s long-context window, the AI guides users through concepts by posing targeted questions rather than delivering long lectures.

Adaptive Feedback: Evaluates user responses in real-time, deciding whether to advance or provide remediation based on the depth of understanding.

Automated Summarization: Uses Gemini 1.5 Flash to culminate each session with a high-speed, personalized summary of key takeaways.

How we built it

We built Mr.So with a focus on modularity and engineering rigor:

iOS Architecture: Developed using SwiftUI and MVVM, utilizing Dependency Injection (DI) to decouple our AI service layer from the UI.

AI Core: Powered by Gemini-3-pro, we leveraged strict JSON Schemas to ensure the model's non-deterministic outputs mapped perfectly to our local data models.

Reactive Logic: Employed Combine to handle asynchronous LLM streams, ensuring the UI remains responsive during complex token generation.

Persistence: Integrated CoreData to allow users to restore their learning state and access their history offline.

Challenges we ran into

The biggest hurdle was managing the State Machine during asynchronous streaming. We encountered complex race conditions where the UI would attempt to transition before the AI had finished its logic. Additionally, ensuring the AI consistently adhered to a "question-first" persona required extensive prompt iteration and the implementation of a validation layer to handle malformed JSON responses. We also spent significant effort polishing the "final question" UX to ensure a clean transition to the course summary.

Accomplishments that we're proud of

YouTube-to-Course Integration: Successfully implemented a feature that parses YouTube video links, allowing the AI to extract and transform video content into structured, interactive Socratic lessons.

Natural Closed-Loop Learning: Created a system that feels like a natural conversation, moving away from static Q&A to a dynamic "mentorship" experience.

Engineering Reliability: Achieved 100% state restoration via CoreData and maintained sub-second response times for streaming AI feedback, ensuring a professional-grade user experience.

Effective Team Delivery: Completed the journey from initial UI concepts to a fully functional, end-to-end AI application within the intensive Google AI Hackathon timeframe.

What we learned

We learned that Prompt Engineering is as much about software engineering as it is about linguistics. Enforcing strict schemas transformed the model from a "chatbot" into a reliable "logic engine". We also gained deep experience in Swift Concurrency and managing complex application lifecycles in a collaborative environment.