MIAX Asesor

It turns the master's material into a knowledge base of 105 topics and about 4,200 sub-concepts. Three specialist agents, for finance, AI and maths, and engineering, debate each question and then write a summary you can download.

RAG over ChromaDB, a FastAPI backend with streaming and a React dashboard.

The master's material is spread across PDFs, notebooks, slides and documents. To study it properly I needed to search all of it at once and get answers from more than one angle, because a question about a model is a maths question, a finance question and an engineering question at the same time.

An extraction pipeline reads PDF, IPYNB, DOCX, PPTX, TXT and Markdown files, with OCR when it has to, splits long texts, structures them into notes per block and routes every concept to a catalogue. The result is 105 topic files with about 4,200 sub-concepts, a ChromaDB index of around 3,069 chunks and an Obsidian vault of 280 files. Gemini and Groq do the bulk extraction for free, with Claude Sonnet as the fallback.

You pick one, two or three specialists: finance, AI and maths, and engineering. With one you get a direct answer grounded in the knowledge base. With two, answers and cross-critiques. With three, answers, critiques, defences and a final synthesis, in about two and a half minutes. The synthesis downloads as Markdown or as a notebook with runnable cells.

The React dashboard streams the debate live and adds a 3D graph of the vault (335 nodes) and a library. Dropping in a new file triggers the whole pipeline, and the agents can use it on the next question. The course material itself stays private; what is shown here is the tool.