Inspiration

We are experiencing a very weird trend in the 2026 developer landscape: the "LLM Crutch". Students are giving up deep comprehension for instant completion, copy-pasting code they don't understand just to get a green checkmark, get past an assignment deadline, etc. We were inspired to build a tool that restores the Socratic Method of learning! Forcing developers to stay in the productive struggle where we believe that real learning actually happens. We were also a victim of LLM, making our logic and skills superficial, but now that we've realised this problem, we're going to solve it ourselves!

What it does

Socratic Eye is a vision-powered AI mentor that watches the workspace in real-time, but not to give you the answer. Using Gemini 3 Flash, our agent sees the image, processes it and then identifies the logic and syntax, intervening only to ask targeted, guiding questions to lead you to your own breakthrough, simulating an AI-powered learning struggle. It helps you in figuring out ways of "how not to do" something, which is where actual learning happens. Every session concludes with a personalised Learning Report that documents the concepts you mastered through failure.

How we built it

We built a system architecture designed for high performance, a cloud native stack powered by Gemini 3 and deployed on Azure.

  • Frontend: A responsive React UI with Socket.IO for low-latency screen streaming without any lag or issues.
  • Backend: A Flask engine utilising Eventlet for asynchronous WebSocket handling and the GenAi pipeline management.
  • Gen AI Core: Gemini 3 Flash (v1alpha) with custom Thinking-Config to maintain the Socratic persona.
  • Database: PostgreSQL for handling authentication and logging user progress and generating growth analytics.

Challenges we ran into

The biggest hurdle which almost made us consider quiting was maintaining a stable WebSocket connection on Azure while streaming high-resolution frame images continuously. We had to deep-dive into Session Affinity (ARR) and server-side "Monkey Patching" to prevent the socket from dropping during intense AI reasoning cycles. We also wrestled with API rate limits, leading us to consolidate our "Interpreter" and "Mentor" agents into a single, efficient multi-modal pipeline.

Accomplishments that we're proud of

We are incredibly proud of our vision-to-mentorship pipeline. Successfully getting an AI to "see" a subtle logic error—like an off-by-one loop error—and then respond with a conceptually accurate hint without spoiling the solution felt like a true breakthrough in AI pedagogy.

What we learned

This project taught us that "User Experience" in education sometimes means adding intentional friction. We learned the intricacies of real-time multimodal AI and how to manage stateful conversations across a distributed cloud architecture using Azure and Docker.

What's next for Socratic Eye!

This is just our first prototype we have plans to make this an actual, scalable GenAI tool. Our roadmap includes:

  • Native IDE Extensions: Socratic Eye in popular IDEs like VS Code as a native extension.
  • Collaborative Struggle: Classrooms where mentors can guide entire study groups simultaneously. This is quite ambitious, and we would love guidance and mentorship to help us accomplish this.

Thank you!

Share this project:

Updates