A deposition-analysis system that catches what humans miss.
3 months to production
Challenge
A legal-tech platform needed to surface answers — and contradictions — buried across thousands of pages of deposition transcripts, fast enough to matter mid-case.
What we built
We built a retrieval system over the transcripts: one model answers questions in context, a second reads for inconsistencies between testimony. Legacy records were migrated out of an aging .NET stack into a clean pipeline.
Outcome
Delivered end-to-end in three months and running on real caseloads.
- AWS Bedrock
- Claude Sonnet — Q&A
- Claude Opus — inconsistency detection
- RAG + ETL