Engagement snapshot
- Mandate
- Speed up trial operations without introducing unsafe or uncited protocol guidance.
- Timeline
- 10 weeks to pilot, 4 more weeks to expand across studies.
- Team shape
- Clinical ops lead, 1 product owner, 2 engineers, and 5 coordinator SMEs.
The problem
Site coordinators lost time hunting through protocol amendments, site manuals, recruitment criteria, and adverse event procedures before they could answer operational questions.
What we built
Built a semantic search and answer system with source-grounded spans, study-arm awareness, and citation views for protocol-sensitive responses.
Operating context
The sponsor was running multiple active studies with frequent protocol updates. Coordination work was becoming slower because teams trusted PDF search less and less as the document base grew.
Key constraints
- Answers had to identify the study, amendment version, and section reference every time.
- No advice could cross studies or arms where inclusion criteria differed.
- The system had to respect role boundaries between coordinators, monitors, and safety staff.
What we built
Study-scoped retrieval
Separated indexing and retrieval by protocol, amendment, and study arm so context stayed bounded.
Evidence-first answers
Rendered concise answers only when evidence windows were strong enough; otherwise defaulted to document navigation.
Operational freshness controls
Watched amendment releases and re-indexed affected documents with visible freshness markers for end users.
Delivery path
Knowledge map
Cataloged which document families mattered for each question class and where teams actually trusted content.
Coordinator pilot
Tested common protocol questions in shadow mode and compared answer quality against coordinator and monitor responses.
Operational hardening
Added version banners, user feedback loops, and admin controls for re-indexing after protocol updates.
Why it mattered
The largest gain came from operational confidence. Coordinators were willing to use the system because it behaved like a fast research layer, not like a fake clinical authority.
Implementation notes
- Protocol-sensitive work benefits from answer restraint, not aggressive answer generation.
- Version labeling is a product requirement in regulated search systems.
- The best evaluation set came from real coordinator questions, not synthetic prompts.