Maverick
An AI coding-assessment platform. It reads a résumé, rates each skill, sets adaptive coding challenges, grades the code and builds a four-week learning path.
A résumé says what someone claims to know. It does not show what they can do today, or what they should learn next.
Five agents, in the order a learner meets them
Profile
Parses the résumé with spaCy, then an LLM rates each skill it finds.
Assessment
Generates coding questions, grades the answers and adjusts difficulty as you go. Submitted code runs in an isolated sandbox.
Recommender
Builds a personal four-week learning path with resources for each week.
Tracker
Notices a stuck learner: three or more days inactive, or three failed assessments in a week. It writes a nudge and can offer an easier warm-up question.
Hackathon
Organisers create hackathons and participants submit code. The agent judges each submission with a score, a plagiarism estimate and written feedback.
What is inside
- Agents
- 5
- Largest agent
- Assessment, 867 lines
- Stuck-learner rule
- 3+ days inactive, or 3+ failed assessments in 7 days
- Tracker API
- 8 endpoints
- Learning path
- 4 weeks
- Leaderboard
- Redis sorted set
Decisions I would defend in an interview
Never run a stranger's code on the server
Every submission executes inside Piston, an isolated sandbox, and only the result comes back.
Difficulty follows the learner
A difficulty adapter raises or lowers the next question from the last result, so the test finds a level instead of assuming one.
Intervene, don't just report
When the Tracker flags someone as stuck it publishes a live alert and generates a warm-up question for that exact skill.
The AI judges, people organise
Hackathons are created by organisers. The AI's job is to score what participants submit.
Built with
- Backend
- FastAPI, PostgreSQL 15, Redis 7, WebSockets
- Models, on Groq
- Qwen3-32B, Llama 3.3 70B, Llama 3.1 8B
- Language tools
- spaCy, Piston sandbox
- Frontend
- React
- Packaging
- Docker Compose