What problem does it solve? Preparing for ML/LLM interviews requires consolidating formulas, derivations, from-scratch implementations, and practice questions into one study document, which is time-consuming to write and error-prone to verify manually. ## Core Features & Use Cases - Structured Cheat Sheet Generation: Produces a 600-1000 line Chinese tutorial with a fixed 12-14 section skeleton covering intuition, core formulas with derivations, from-scratch PyTorch code, variants, complexity analysis, and 25 interview questions tiered as L1/L2/L3. - Cross-Model Review Loop: Submits drafts to a gpt-5.5 reviewer (fresh thread, xhigh reasoning) running 10 checks on math correctness, code executability, citations, and style rules, iterating until PASS. - HTML Rendering with Audit Trail: Renders the Markdown via render_html.py with academic template and merges math/code and render review histories into a single review.json audit file. - Use Case: Ask for a cheat sheet on "RLHF / DPO / PPO" and receive a reviewed Chinese tutorial MD plus rendered HTML, with no auto-commit so you review before publishing. ## Quick Start Ask the assistant to write an interview cheat sheet on a specific ML topic such as "MoE (Mixture-of-Experts)" with balanced effort.
