TorchLeet

Why we included this project

For engineers prepping for machine-learning interviews, TorchLeet is one of the more honest training resources around. It collects 65 PyTorch coding problems sourced from real interviews at Google, Meta, Anthropic, OpenAI, and similar companies, each packaged as a Jupyter notebook with a fill-in-the-blank question sheet and a separate solution notebook. The coverage is current: alongside basics like attention blocks and normalization layers, it walks through LoRA, DPO, PPO for RLHF, GRPO, KV caching, speculative decoding, and continuous batching, so you practice the techniques modern ML interviews tend to probe. An auto-grader checks your implementations, and the optional MCP-based AI tutor coaches with progressive hints rather than handing over answers. It works best as a deliberate-practice lab for engineers who already know PyTorch and want to rebuild core concepts from scratch, not as a quick reference.