The Challenge: Data Limitations Slowing AI Engineers
AI engineers frequently face three key issues:
Data Scarcity: Incomplete or noisy data limits the accuracy of models.
Manual Preprocessing: The process of cleaning and organizing datasets is labor-intensive and time-consuming.
Disconnected Expertise: It’s difficult to translate domain knowledge into models, leading to fragmented and slow communication with experts.
These challenges resulted in inefficient workflows, making it difficult for engineers to meet the demands of predictive maintenance and similar applications.
Approach: Centering AI Development on Expertise
To overcome these limitations, K1st BUILD was designed to enable AI engineers to build models by directly capturing and applying expert knowledge rather than relying solely on vast datasets. Working closely with engineers like Zhang Yu-san from CNA (China & Northeast Asia Company by Panasonic Group), I explored their workflows and identified the key pain points.
The key components of this approach included:
Knowledge Capture: Engineers could input domain-specific insights, bypassing data-heavy processes and improving model accuracy. This shifted the focus from cleaning data to using real-world expertise.
Knowledge-to-Model Translation: K1st BUILD facilitated the conversion of expert knowledge into AI models through symbolic logic, reducing reliance on raw data while ensuring nuanced insights were reflected in the models.
Dynamic Model Updates: Engineers could continuously refine models as new knowledge became available, ensuring adaptability without starting from scratch.
By addressing these pain points, K1st BUILD allowed engineers to build models faster and more accurately. Before, they struggled with delays due to data shortages and disjointed communications with domain experts. After, the system enabled real-time integration of expert knowledge, reducing development time and improving model precision.
Prototype Feedback and Iterative Refinement
Early prototypes of K1st BUILD were tested with both engineers and domain experts. Engineers found the UI intuitive and appreciated how easily they could incorporate expert insights into models without manual intervention. Domain experts reported that the tool allowed them to provide real-time feedback, streamlining collaboration between them and the engineering teams.
Working closely with front-end engineers, I ensured a smooth handoff by developing a clear design checklist. Within three months of joining Aitomatic, I created a design system for K1st BUILD.
Impact: Faster, More Accurate Model Development
Before K1st BUILD, model-building was often slow and lacked the precision needed for critical applications. After implementing K1st BUILD, engineers gained access to a system that allowed them to integrate expert knowledge in real time, leading to faster development and more accurate models.
30% Faster Model Creation: Engineers reduced time spent on data preparation by directly incorporating expert knowledge.
15% Increase in Model Accuracy: By leveraging domain-specific insights, models became more accurate, leading to better predictions and outcomes.
Enhanced Collaboration: The tool bridged the gap between engineers and domain experts, fostering a smoother and more efficient collaboration.
Reflection
K1st BUILD was about letting experts drive the AI development process, not just relying on data. It was refreshing to see how much value we could get from human expertise instead of just more data.
The challenge was in designing a system that made it easy for engineers to input and apply expert knowledge without slowing them down. The biggest lesson for me was that designing for human expertise means building tools that work the way people think, not just the way machines process data.