The research repository that became the team's AI agent

• Problem: research was used once, in the meeting it was presented in, then forgotten.
• What I did: made past research searchable in plain language, put findings in front of people every week, and cut the time from session to evidence.
• Result: about 30 people a week used the repository on their own, and its infrastructure became the base of the team's UXR agent, which now joins research with product telemetry and sales data.

A study ended in a read-out deck. The deck was shared once, and the finding lived on only in the heads of the people in the room. Six months later a product manager would ask a question the team had already answered, and either nobody remembered, or the answer was in a file nobody could find. So research was paid for twice, or the decision was made without it.

The problem wasn't that research was missing. Finding it cost more effort than asking again, or than guessing.

A repository you can ask questions. I designed and built an org-wide research repository that now holds more than 88 studies. You search it in plain language, such as "what do analysts say about alert fatigue", and it returns the studies and findings that answer the question, each linked back to its source. I built it on AWS Lambda and DynamoDB, inside the company's own cloud.

A newsletter every week. Each issue put recent findings in front of people, with links into the repository, so research reached people who never came to a read-out.

Evidence in days, not weeks. Sessions were transcribed with AI and cut into short highlight reels, so stakeholders heard participants in their own words within days of a session. Read-outs became weekly instead of one at the end of a study, so the team could act while the study was still running. Write-ups moved from slides to exhibit-style pages: a claim, the evidence for it, and what it changes, in that order.

The repository's infrastructure became the foundation of the team's UXR agent, built by a senior colleague. The agent joins interview transcripts with product telemetry and sales data, so a question can be answered from what people said, what they did and what they bought, in one place.

That's the outcome I'd point to first. A repository is only worth something if people keep using it. Having someone more senior build the next tool on top of it is a stronger sign than any visitor count.

• Plain-language search over tags. Tagging schemes decay as soon as the person who designed them stops maintaining them. Search on the findings themselves keeps working.
• Every answer links to its source. A summary without a path back to the study is just an opinion, and people stop trusting it the first time it's wrong.

The repository averaged about 30 unique visitors a week, used by people looking for answers on their own, not only by researchers. The newsletter went out every week. Evidence reached stakeholders in days through highlight reels and weekly read-outs. The infrastructure became the base of the team's UXR agent.

Measure what matters more than visits: how often a search answers the question without a new study, and which teams never search at all.