TorchLeet
A collection of 68 PyTorch problems drawn from real machine learning and AI engineering interviews at companies such as Google, Meta and Anthropic, worked through in Jupyter notebooks with an optional MCP-based AI tutor that gives progressive hints without the answer.
TorchLeet is a set of 68 PyTorch practice problems taken from real machine-learning and AI engineering interviews at companies including Google, Meta and Anthropic, organized into a Basics track, an LLM-building track that constructs a language model from scratch step by step, and an Advanced track covering systems, kernels, modern architectures and alignment; some questions are tagged with the company that asked them.
Each problem is a Jupyter notebook with a question file and a separate solution file, meant to be solved by hand rather than by pasting into a general chatbot, since the intended learning happens in struggling through the problem.
An optional MCP server turns a connected AI assistant into a tutor with access to all the problems, offering progressive hints, company-specific mock interview prep, a senior-engineer-style code review, and a no-spoilers teaching style, through four separate guides for tutoring, interview practice, review and deep explanation.