The problem
Keyword searches and fixed filters still left consultants reviewing many device records by hand. Much of the useful detail was inside PDF documents rather than readily searchable fields.
What we built
We built an LLM-enhanced search implementation that helped compare the language describing a device’s indications for use.
- Made document information usable for search.
- Supported queries describing the device and its intended use.
- Used semantic matching to find relevant device records.
What changed
More relevant records. A better starting point for review.
A consultant could search for a specific description such as a non-invasive maternal fetal heart monitor and retrieve relevant records instead of beginning with a broad keyword query. The tool supported the consultant’s research; it did not make the regulatory decision.
The technical decisions
Why use AI here?
The task depended on comparing the meaning of language, not just matching an exact product term. That gave the model a specific job within the search system.
Keep the expert in the process.
Search results were a way into the underlying device records. Consultants still needed to examine the evidence and determine its relevance to their submission.