LeetCoach

Project Thumbnail

LeetCoach was inspired by the need for a more innovative, more interactive way to practice algorithms and data structures. I aimed to develop an AI-powered coding coach that not only grades answers but also actively guides developers with real-time feedback, strategic hints, and complexity analysis. Using OpenAI’s GPT-OSS-20B model, I built a system that combines privacy, speed, and coaching power, running entirely locally with CUDA acceleration.

Throughout this project, I learned how to integrate large language models with GPU acceleration for optimal performance, design a scalable microservices architecture, and implement real-time code execution and judging systems. I also deepened my experience in building modern interactive web interfaces that create an engaging, professional coding environment.

The platform itself was built with a FastAPI backend in Python, paired with PostgreSQL for persistent storage. AI inference runs through GPT-OSS-20B, accelerated with CUDA for fast response times. On the frontend, I used Next.js with TypeScript and Tailwind CSS to deliver a responsive, smooth user experience. The infrastructure is containerized with Docker and orchestrated with Docker Compose, while the custom judging engine provides secure, sandboxed code execution.

One of the biggest challenges was optimizing inference speed with CUDA to ensure real-time coaching without lag. Another was building a secure environment for running user-submitted code, which required careful isolation and resource management. Designing an intuitive problem-generation system with multiple difficulty levels also pushed me to balance complexity with usability. Ultimately, the most significant challenge was ensuring that all of these moving parts came together seamlessly to deliver an interactive, scalable platform that feels natural for developers preparing for technical interviews.