Design-partner study. Anonymized at partner's request.
The problem
A fraud investigations team with 12,000 case writeups. Patterns repeat — when a new case comes in, the first 90 minutes are spent finding similar prior cases to inform the investigation. The team had been using a wiki with full-text search, but: (1) full-text missed cases that used different vocabulary, (2) there was no history (when a case got revised, the prior version was lost), (3) PDF exports for legal had to be regenerated manually each time.
What changed
- Hybrid search (BM25 + embedding) finds related cases across vocabulary differences. "Synthetic ID rings" matches "fabricated identity networks" if the embeddings agree.
- Every case writeup is versioned bitemporally. "What did this case say on the day it went to court?" is a query, not an archeology project.
- Compliance bundle export is one click. Legal gets PDF + JSON + signed manifest.
- AI assistant trained on the team's writing style (voice profile) drafts initial case summaries; investigators edit instead of writing from blank.
Result
- First-90-minutes-of-investigation: from 90 min to 18 min
- Cases reviewed per investigator per week: +340%
- Legal-export turnaround: same day; previously 2–3 days