Legal-tech patent intelligence platform, 2025
LLM deep search for patent infringement
I led a legal-tech platform that uses LLM deep search over thousands of patents to flag potential infringement, with evals for accuracy, false positives and explainability built for legal review.
The system
- Patent corpus
- Embeddings
- Multi-stage retrieval
- LLM analysis
- Legal review
Disciplines I coordinated
Engineering, ML & evals, compliance
- Thousandsof patents compared for potential infringement
- Evalsfor accuracy, false positives and explainability
Context
Checking a patent for potential infringement means comparing it against thousands of others, written in long, technical legal language across jurisdictions. Any flag the system raises has to hold up in front of a legal team.
My role
I led the product from discovery to architecture design.
What I did
- Deep search, not keyword search. Wrote the workflows and PRDs for LLM search built on embeddings, semantic similarity and multi-stage retrieval (RAG), so each patent is compared on meaning across thousands of others.
- Evals before launch. Defined the evaluation criteria with the MLOps team: accuracy thresholds, false-positive reduction and explainability, so a lawyer can see why something was flagged.
- Built for legal workflows. Audit logs that match how legal teams review and defend their work.
- Document processing. Extraction, classification and summarization for long technical legal documents across jurisdictions.
Outcome
Led from discovery through architecture design, with the evaluation framework defined before build.
Stack and tools
- LLMs
- Embeddings
- Vector search
- RAG
- Evaluation framework