one mastered record, every field traced to its sourceVisuals recreated — no client data
The Problem
Clinical study intelligence at Trinity lived everywhere and nowhere: 847K+ studies spread across ClinicalTrials.gov, Citeline, Cortellis, and internal sources, accessed through scattered spreadsheets and ad-hoc queries. Every question about the study landscape meant hunting across systems, reconciling conflicting records, and trusting whichever export was most recent.
Master data platforms are traditionally built as dense administrative consoles — powerful for data stewards, impenetrable for everyone else. The team needed the opposite: a way for anyone to interrogate the unified corpus directly.
In a regulated domain, an answer without a source is a liability. The citation layer couldn’t be a footnote — it had to be the product.
The Approach
I designed the platform around a conversational interface as the primary way in. Instead of navigating hierarchies of tables and filters, users ask questions in natural language and the system resolves them against the unified corpus.
The hard design problem wasn’t the chat — it was trust. Pharma and life-science teams can’t act on an answer they can’t verify. So source attribution became a first-class design surface, rendered inline as cards in the conversation itself.
And where the system acts on the data — proposing which records across sources are duplicates — I kept a human in the loop by design. The AI recommends and explains its reasoning; a data steward makes the call. Trust wasn’t only about citing an answer; it was about keeping a person in control of every automated decision.
On evidence: Trinity is a life-sciences consultancy, so people who interrogate study landscapes for a living sit in the building. They were my proxy for the end user — I designed against internal domain experts rather than researchers at client organisations. That’s a real limit. An SME can tell you whether an answer is correct; they can’t tell you whether someone under deadline will trust it enough to skip opening the source. The closing section names what I’d put in front of real users first.
Evidence base — internal SME review · domain-expert working sessions · stakeholder walkthroughs. No direct end-user testing. The days-to-hours research time is reported by the product team and its users, not instrumented.
1 / 4Ask, verify, decide
Key Decisions
Chat as the front door to master data
A conversational interface replaced the spreadsheet-and-query workflow, letting users interrogate 847K+ studies without knowing where each record lives or how the sources differ.
Source attribution rendered inline as cards
Every AI response carries its provenance — citation layers rendered as inline cards in the conversation, so verification happens in context instead of in a separate audit trail.
Designed for multi-source reconciliation
ClinicalTrials.gov, Citeline, Cortellis, and internal data don’t agree with each other. The interface makes the origin of each fact visible field by field — users understand not just the answer, but which system it came from and where the sources disagree.
Human-in-the-loop match review
Beyond answering questions, the platform resolves duplicate studies across sources. The AI recommends match pairs with its reasoning and its evidence; stewards confirm, reject, or defer. The person is always the final arbiter — the AI’s job is to make the decision faster and better-informed, never to make it alone.
AI analysis that explains, not just answers
On a comparison or a record, the AI surfaces a short rationale with supporting and contradicting signals and a confidence level — so a reviewer can judge the model’s thinking instead of trusting a black box.
Outcome
Days → hoursstudy research cut from multiple days to a few hours by consolidating source review, matching, and AI-assisted analysis into one workflow — reported by the product team and its users, not instrumented
847K+clinical studies unified behind a single conversational interface
4 → 1data sources reconciled into one queryable corpus
The citation and explainability patterns designed here were generalized into the Trinity Design System’s AI module — becoming the standard for how Trinity’s AI products earn user trust.
What I’d Validate Next
The citation layer is the whole thesis of this product, and it’s the part I have the least direct evidence for. Three things I’d test with real users before treating it as settled.
Does anyone actually open the citation?
The premise is that inline provenance earns trust. If instrumentation showed citation cards are almost never expanded, then trust is coming from somewhere else — the interface’s confidence, the brand, or nothing at all — and I’d be investing in the wrong surface.
Chat versus filters, for the expert
Natural language clearly helps a newcomer. A data steward who already knows the schema may find it slower than a filter they can aim precisely. I’d run both against identical tasks with both audiences before committing further to chat as the only front door.
What happens when the sources disagree
Showing the origin of each fact assumes the user can adjudicate between them. I’d watch real reconciliation tasks to find out whether the interface resolves the conflict or simply hands it to the person — which would make it an honest UI and a failed one.
