www.github.com/d-pap/codecoach
Purpose
LeetCode is great for interviews but weak on ICPC problems and on-the-spot feedback. My ACM club wanted one place where students could solve past contest questions, run code in the browser, and get real help from an AI mentor. So three of us spent eight months turning that wish list into Code Coach.
Outcomes
200+ users across two semesters; average weekly return rate 48 %.
3000+ code runs with language-level timing and memory stats stored for leaderboards.
15% faster solve rate after we shipped the AI hint system (club survey, n = 41).
AI-powered Resume builder used by 62 students; exported 120 tailored resume PDFs.
Features
Problem sets – ICPC archives plus LeetCode-style practice.
In-browser compiler – Python, C++, Java; full Judge0 output in < 5 s.
AI assistant – context-aware hints, step-by-step reasoning, gated full answers.
Resume builder – one click from form data to job-specific PDF using the same LLM.
Progress tracking – execution history, success rates, best runtimes.
Tooling
Frontend. React with MUI components and React Query for caching. I designed every screen and wired all state flows.
Authentication. AWS Cognito handles email and social sign-in and controls access to compute-intensive features.
API layer. AWS API Gateway routes to Node.js Lambdas. Payloads are typed, requests are throttled, and functions are sized for cold-start budgets.
Data. MongoDB Atlas stores problems, submissions, and per-user AI token counts.
Compiler sandbox. Judge0 CE is called from Lambda to run code safely in multiple languages.
AI assistant. GPT-4o served through UM-GPT (University of Michigan instance). Fine-tuned on ICPC Q&A with prompt rules that force hint-first answers.
Dev-ops extras. GitHub Actions for CI, CloudWatch for logs, and React Testing Library for component tests round out the stack and keep quality high.
Behind the scenes
