Show HN: Md-pdf-md – Bidirectional Markdown<>PDF with local AI vision (www.npmjs.com)

md-pdf-md is an open-source, zero‑configuration CLI and library that converts Markdown to polished PDFs and — crucially — extracts structured Markdown from PDFs using local AI vision. It bundles themes (GitHub, GitHub Dark, Academic, Minimal), smart page breaks, auto TOC with page numbers, Shiki-powered syntax highlighting and Puppeteer-based PDF rendering. The standout feature is bidirectional conversion: MD→PDF is fast (2–3s for typical docs), while PDF→MD uses a vision-language model (LLaVA) run locally via Ollama to identify headings, lists, code blocks, tables and even programming languages, claiming ~90%+ structure preservation. For AI/ML practitioners and developers this matters because it brings reliable, private PDF understanding to the desktop — no cloud APIs, no data leak — enabling editable workflows (extract→edit→rebuild) and programmatic document processing. Technical requirements: Node.js ≥16, Ollama for PDF→MD, and pulling the LLaVA model (~4.7GB). Performance is ~5–10s per page on CPU (2–5s on GPU). The project is MIT‑licensed, free, and positions itself as a more structure-aware alternative to traditional tools like pandoc for PDF extraction, making it useful for reproducible docs, research pipelines, and secure enterprise workflows.