Client

Context

Developing an AI-driven image search function for a jewelry webshop to improve the time-consuming task of manually classifying jewelry.

Imagine being presented with a box of 1000 rings, each with subtle differences that make classification difficult and time-consuming. This was the reality for our client, which made an AI solution necessary to streamline and automate the classification process.

01 Define

Start with why

The first step was to challenge the concept of classical classification. We identified the limitations of retraining the AI model for every new product and the inefficiency of classifying extensive classes. This led us to propose an alternative approach, similarity embedding, which compared products based on visual similarity. This machine learning method allows for accurate classification of extensive classes without the need for retraining the AI model.

User interviews

Research

Brainstorming

02 Ideate

Generate ideas and concepts

Through technical analysis and co-creation with the client, we focused on the specifics of implementing the similarity embedding approach. In order to completely understand the client's needs and preferences, we worked closely with them to iteratively improve the design document. We made sure that the solution matched their objectives and was in line with their vision by utilizing the expertise of our team and actively involved the client.

Service design

Workshops

Technology exploration

03 Build

Building the solution

The development of the AI-driven image search function was executed by our sister company, Brainjar. Our role was to facilitate the development process, coordinating with the development team, ensuring seamless communication, and managing the budgetary aspects. In order to handle any concerns or remarks quickly, we kept the client informed and involved throughout the build phase.

Development

Product ownership

Project management

04 Grow

Nurturing and growing

Analytics

Project management

Results

At the end of the project, the AI-driven image search function delivered impressive results. The search speed was reduced to 5 seconds, surpassing the maximum threshold of 10 seconds set in the non-functional requirements. Additionally, the accuracy of the model reached 93%, exceeding the target of 90%. This achievement demonstrated the effectiveness of the solution and its potential for further improvement as the AI algorithm continues to learn and optimize over time.

Model's accuracy: 93%

Search speed: 5 seconds

As a technology agnostic company, we always explore multiple ideas, technologies, and even solutions. We map your processes in a detailed manner so the solution is always tailored to your needs and remains future-proof. However, as we're just human beings, we have a tool stack we often use and have become proficient in using.

Technologies

Take a look at the tool stack we use

Discover other work

Product Ownership

User Testing

Brainhouse - Building a web browser for ALS patients

All cases