| name | codegym-concept |
| description | Product concept and vision for CodeGym — an AI-powered coding practice platform that goes beyond DSA. Use this skill when making product decisions, evaluating feature scope, or ensuring work aligns with the core vision. This is the "why" and "what" document. |
CodeGym — Product Concept
One-Liner
LeetCode trains you to pass interviews. CodeGym trains you to be an engineer.
Core Thesis
The bottleneck to learning a new library, pattern, or tool is not documentation — it's the absence of a tight feedback loop. CodeGym closes that loop by AI-generating sandboxed, test-verified coding problems from a single natural language prompt across any engineering domain.
Target Users
- Developers learning a new language, framework, or tool
- Engineers strengthening weak areas ("I've never written raw SQL window functions")
- Teams onboarding engineers to internal patterns
Three Pillars
1. AI Problem Generator (Coding Mode)
User enters a prompt like "practice Express pagination" → Gemini generates:
- Problem statement (Markdown)
- Test cases (hidden, used for grading)
- Full solution (agent-validated)
- Skeleton code (solution with logic stripped — user fills it in)
The key innovation: An agent writes the full implementation first, validates it against generated tests, then strips the logic to create the learner-facing skeleton. This inversion guarantees solvability.
2. Adaptive MCQ Marathon
- User enters a free-text prompt describing what they want to practice (e.g. "React hooks", "SQL window functions")
- Onboarding phase: AI generates 3-8 conversational skill-assessment questions to calibrate difficulty
- Onboarding answers saved as notes in the memory system (persists across sessions)
- MCQ questions generated using prompt + onboarding notes + existing memory for maximum personalization
- "I'm Feeling Lucky" skips onboarding — generates questions using existing memory only (fast path for returning users)
- Difficulty adjusts based on rolling performance (last 10 answers) + self-assessed level
- Memory system has two complementary fields:
skill_summary (AI narrative) and notes (raw Q&A history)
- Good for rapid knowledge assessment before diving into coding problems
3. Skill Memory & Profile
- Persistent memory with two complementary fields:
skill_summary — AI-generated narrative of strengths, weaknesses, and learning trajectory (updated by Gemini after every interaction)notes — Raw onboarding Q&A history (what the user told us about themselves, appended over time)
- Updated after every MCQ set and coding submission
- Onboarding answers from each marathon session accumulate in notes, giving the AI richer context over time
- Generates a natural-language "skill report" — who you are as an engineer, what you're versed in, what you struggle with
- Powers adaptive difficulty across both modes
Problem Domains (Full Vision)
- DSA — Standard competitive programming
- API Design — REST, GraphQL, gRPC, WebSockets, Auth
- Frameworks — Express, FastAPI, Spring Boot, Django, etc.
- Design Patterns — GoF, SOLID, refactoring exercises
- Databases — SQL window functions, CTEs, schema design, query optimization
- Concurrency — goroutines, asyncio, tokio, synchronization primitives
- Systems — malloc, file I/O, interpreters, bit manipulation
- DevOps — Docker, Bash, CI/CD, Nginx, Terraform, K8s
- Frontend — DOM APIs, state management, Web performance
- Security — HMAC, encryption, input validation, rate limiting
- Messaging — Kafka, RabbitMQ, event buses, outbox pattern
- Networking — TCP sockets, HTTP parsing, DNS, TLS
Hackathon Scope (March 21, 2026)
Stack: Next.js 15 (Bun runtime) on Vercel + Gemini 3.1 Pro + Supabase + BetterAuth + ElevenLabs + Sentry
MVP features (must ship):
- AI Problem Generator with Monaco editor + in-browser execution
- Adaptive MCQ marathons
- Skill profile dashboard
Cut for hackathon:
- Docker execution engines / Judge0
- Multi-step projects
- Team/org mode
- Languages beyond JS/TS/Python (for execution)
Demo Script (3 minutes)
- Open CodeGym → "Generate a problem about Express middleware"
- Watch Gemini generate a full problem in real-time (streaming)
- Solve it in the Monaco editor → run tests → see pass/fail
- Switch to MCQ Marathon → type "React hooks" → answer 3-5 onboarding questions about your React experience
- Watch Gemini generate personalized MCQs calibrated to your assessed level
- (Or click "I'm Feeling Lucky" to skip onboarding and go straight to questions)
- Open skill profile → see updated strengths/weaknesses + onboarding notes in memory
Anti-Patterns (What CodeGym Is NOT)
- Not a basic RAG chatbot
- Not a generic "AI tutor" or education chatbot (banned by hackathon rules)
- Not an interview prep platform — it's an engineering skills platform
- Not Streamlit (banned)
- Not medical/mental health/nutrition (banned)
Competitive Landscape
| Platform | Focus | AI-Generated? | Adaptive? | Beyond DSA? |
|---|
| LeetCode | DSA interview prep | No | No | Limited |
| HackerRank | DSA + some domains | No | No | Some |
| CodeCrafters | Systems projects | No | No | Yes (narrow) |
| Exercism | Language tracks | No | No | Yes (narrow) |
| CodeGym | All engineering | Yes | Yes | Yes |
