Multimodal Research Agent & Evaluation Platform

A personal research tool that ingests scientific PDFs across radar and computer vision literature and makes them genuinely searchable — not just the text, but the tables and figures too.

What it does

  • Processes scientific PDFs with Nemotron Parse, preserving document structure across text, tables, and figures
  • Indexes content with embedding models and vector search, with reranking for precision
  • Runs agentic workflows for evidence gathering, cross-paper comparison, and grounded summaries
  • Evaluates itself: retrieval quality, faithfulness, citation correctness, and end-to-end task success are all measured, not assumed

Stack: Python, embedding models, vector search, rerankers, LLM agentic workflows, Nemotron Parse