Semantic Search

During my time at CareerBuilder, I had the priviledge to work on Semantic Search, a feature based on AI and Machine Learning to help recruiters search more efficiently.

Why this project is special to me

I got to work some with of the most talented Machine Learning engineers.

My Roles

  • Worked with UX Researcher and PM to iterate and test
  • Created interactive prototype
  • Provided guidance and worked closely with engineers to build and test prototype

Search Personas

Some examples of key persona

The Semantic Search tool was built on top of CareerBuilder’s Resume Database and both were key projects involving big investments for the company. Our team was tasked to get this product out to the our customers: the recruiters. We wanted to see the value behind the AI powered search engine by letting our users interact with it.

Our product team consisted of engineers, a UX researcher, product manager, executive stakeholders and a designer. This team was aligned by continous rounds of user interviews and testing. As a result, we were able to continously test, get feedback and iterate. The feedback we got from users were the results of real-time search prototypes developed by the engineers, as well as interactive prototypes I created.

Established search habit was tough to change

Half of our users are seasoned recruiters with established habits who are using the boolean operators. They expected the search engine to work without having to change their workflow. As a result, we had to accomodate the search to work on both boolean and semantic operators.

Semantic Search Wireframe

Example of wireframe iteration we showed during our testing

Forming the search criteria using Semantic Search

Through testing we learned that users would love the search suggestions to happen as they’re typing in the search form. We adjusted the semantic search engine to show results, while user was forming of search criteria.

Search suggest

This iteration shows search suggestions under the search box

“Just show me the results”

In the beginning we leaned into exposing more controls and interactions. Over time, we learned that users expected the results to reflect the “smartness” of the search engine. They didn’t care as much about granular control and they expected to see the relevant results reflected. With this knowledge, we opted for simpler interactions.

This search iterations shows granular control of keyword organizations.

This iteration shows granular control of keyword organizations.