Operational Intelligence
Contract Risk Radar
Clause-level contract review grounded in past SME verdicts, with a rationale for every flag.
About the Project
A three-stage pipeline (obligation extraction, per-obligation judgment, verdict prediction by analogy) reviews contract clauses against obligations drawn from three policy documents. Each flagged clause gets a verdict (Accept, Modify, Replace, Reject or Escalate), a confidence flag and a rationale, and low-similarity clauses default to Escalate. Built and evaluated on synthetic contracts.
Technical Overview
Data Sources
- Cybersecurity contract clauses (Word files)
- Policy documents (Word files)
- Clause-level SME review verdicts and comments
- Obligation catalogue (133 items)
ML / NLP Stack
- LLM structured extraction with Pydantic schemas
- Embedding similarity retrieval (nomic-embed-text)
- Analogy-based verdict prediction
- Batched LLM judgment (10 obligations per call)
- Drafted replacement clause text (gated)
Key Challenges
- Clauses with no close precedent in the example set
- Overlapping policy obligations from three source documents
- Only 38 labelled SME examples to learn from
- Choosing between escalate and modify when few close precedents exist
Interested in this solution?
Let's discuss how we can adapt this approach for your business.
Get in Touch