I need a complete workflow that lets me drop a .docx exam paper into the system and receive a clean, separate answer sheet in return. The papers follow a strictly Multiple choice structure and often include embedded images next to the options, so the agent must preserve those visuals when it splits the document into individual questions. Scope of content • Subjects the prototype must handle first: – Mathematics – Science Core tasks 1. Parse the incoming .docx, detect every question boundary, pull out the text, any inline or anchored images, and the A-D options exactly as they appear. 2. Run an AI reasoning layer that selects the correct option for each question. (I am open to rule-based filtering plus a large-language-model, or any hybrid that gets reliable accuracy.) 3. Produce a concise summary document—Word, PDF, or Markdown is fine—that lists Q-number and the chosen answer only. No edits should be made to the original file. Technical notes I expect Python for the pipeline (python-docx, Pillow/Tesseract for image handling, and whichever LLM API you prefer). Make sure the code is modular so I can later extend it to other subjects. Acceptance criteria • Drop-in demo with at least two sample papers (one Maths, one Science) showing >80 % correct answers. • Output summary file generated automatically, named <original_filename>_answers. • README with setup steps and clear instructions on where to place new papers for processing. Hand over the runnable code, the sample papers you used for testing, and the generated answer sheets.