I classified 3.5M US patents with Nemotron 9B on a single RTX 5090 — then built a free search engine on top

A patent attorney has launched PatentLLM.org, a specialized search engine covering 3.5 million US patents from 2016-2025. The entire system was built using local hardware: an NVIDIA RTX 5090 GPU ran NVIDIA's Nemotron 9B model for 48 hours to classify all patents into 100 technology categories. The data resides in a single 74GB SQLite file with FTS5 (Full-Text Search) indexing, enabling sub-second query responses without external dependencies.

Unlike typical AI-powered search that uses vector embeddings for semantic similarity, this system prioritizes exact phrase matching crucial for patent law. When an attorney searches for "solid-state battery electrolyte," they get documents containing those exact terms rather than semantically similar concepts about "energy storage." The search uses BM25 ranking with custom weights (title: 10.0, assignee: 5.0, abstract: 3.0, claims: 1.0) and includes natural language query expansion via local LLM before converting to FTS5 boolean queries.

The entire stack demonstrates efficient local AI deployment: FastAPI + Jinja2 serves the interface, while a Chromebook handles hosting via Cloudflare Tunnel. This approach eliminates API costs and privacy concerns while delivering professional-grade search performance. The project showcases how specialized domain knowledge (patent law requirements) combined with accessible AI models can create powerful tools that outperform generic solutions for specific professional use cases.