Case study · Puxxle
Every answer carries its evidence.
An evidence-first research backend strips PII, indexes transcripts, and cites every answer before it reaches review.
- Sector
- Research
- Function
- Product AI
The operation
Research review depends on being able to inspect the source behind an answer. Transcripts also bring personal information that should not travel into the working evidence set.
What AI should own
The backend should strip PII, index transcripts, and retrieve the evidence for an answer. The citation remains attached so review starts from the source, not from trust in a generated sentence.
What shipped
An evidence-first research backend: PII stripped, transcripts indexed, every answer cited.
Reconstructed answer trace
Evidence decides whether an answer exists.
This trace reconstructs the approved backend rule: PII is stripped, transcripts are indexed, and every answer is cited before review.
Illustrative reconstruction, not a Puxxle screen or transcript.
- 01Transcript
Input to the research backend
- 02PII stripped
Personal information removed
- 03Evidence indexed
Transcripts made retrievable
Is supporting evidence available?
Cited answer → Review
No supported answer
The numbers
The engagement measured a 70% reduction in review time. The approved record also describes the answers as 100% citation-anchored.
- 70%
- less review time
- 100%
- citation-anchored answers
What runs today
The research backend prepares cited answers for review. Evidence stays inspectable, and personal information is removed before transcripts enter the indexed source set.