autocurricula 0.1.0

Adaptive coding interview practice powered by Claude

autocurricula

Adaptive interview practice powered by Claude. Problems are generated on-the-fly based on your role and skill level, with real-time feedback and structured test results. Covers coding, math, probability, statistics, algorithms, and brainteasers.

Requirements

• Python 3.11+
• Claude CLI installed and authenticated

Getting Started

pip install autocurricula
autocurricula

On first launch, create a workspace by choosing a role (e.g. "ML Engineer", "Quant Researcher", "Backend Developer"). Claude generates problems tailored to that role and adapts difficulty as you progress.

On first run, autocurricula creates a sandboxed virtual environment at ~/.autocurricula/.sandbox_venv with numpy, pandas, scipy, torch, and pytest for executing solutions.

Features

Adaptive Problem Generation

Problems are generated by Claude based on your role, category, and current difficulty level. The system tracks your solve rate and self-rated difficulty to adjust what comes next. Categories rotate to ensure broad coverage.

Problem types:

• Code problems -- write a function, run open tests, then submit for hidden tests and Claude review
• Derivation problems -- written answers for math proofs, probability puzzles, system design reasoning, and conceptual questions

Categories: Math, Probability, Coding, Statistics, Algorithms, Brainteasers

Difficulty levels: Easy, Medium, Hard (auto-adjusted based on performance)

Code Editor

Monaco editor with Python syntax highlighting, autocompletion, hover documentation, and function signatures powered by Jedi. Supports numpy, pandas, scipy, and torch out of the box.

Test Runner

Each code problem comes with an open test suite (visible while solving) and a hidden test suite (run on submit). Tests execute in an isolated sandbox with a 30-second timeout. Results are displayed with structured pass/fail/error status and expandable failure details showing expected vs. actual output.

Claude Review

When you submit a solution, Claude reviews your code alongside test results and gives a verdict: solved, retry, or move on. Feedback explains what worked, what didn't, and what to try next.

Scaffolding

Stuck on a problem? Request a scaffold and Claude generates an easier prerequisite that targets the specific concept you're missing. Solve the prerequisite, then return to the original problem.

Theory

Each problem comes with a theory document covering relevant background: formulas, derivations, algorithmic intuitions, and worked examples. Rendered with LaTeX support.

Chat

Ask Claude for hints and clarifications without getting the answer. Claude sees your current code and the problem statement for context. Chat history is saved per problem.

Progress Tracking

Track your progress across categories with solve counts, attempt counts, and success rates. Rate each solved problem's difficulty (1-5) to help the system calibrate future problems. Problems you solved but rated as hard are flagged as "high-regret" for potential revisiting.

Workspaces

Maintain separate workspaces for different roles. Each workspace has its own problem history, progress state, and difficulty curve. Switch between workspaces from the landing page.

License

MIT. Free to use, modify, and distribute for personal and commercial purposes. See LICENSE for details.