LegalTech
Lumen Legal — AI document review platform
Overview:
Lumen Legal, a contract-review SaaS, needed an AI layer that could extract clauses, flag risks, and suggest redlines across 100+ contract types. Accuracy and citation traceability were non-negotiable — lawyers can't trust unverified output.
Approach:
Helios built a RAG pipeline using Anthropic Claude with citation enforcement, custom prompt templates per contract type, and a golden eval set of 2000 labeled clauses. Every extraction surfaces the source paragraph; lawyers can accept or override with one click. Eval runs nightly with regression alerts.
Engagement details:
Duration: 10 weeks prototype to production
Stack: Next.js, Anthropic Claude, pgvector, LangSmith, AWS
Team size: 4 (1 architect, 2 AI/full-stack, 1 prompt engineer)
Eval: 2000-clause golden set with weekly regression runs
Outcome:
92% accuracy on the golden set at launch (vs 76% human baseline at same speed). 8x faster end-to-end review cycle. 4 mid-size law firms onboarded in first quarter post-launch.
92%
Extraction accuracy (golden set)
8x
Faster contract review
4
Law firms onboarded
