K1st BUILD

Build AI Models from Domain Knowledge

Company

Aitomatic

Role

Product Designer

Year

2022 – 2023

Overview

K1st BUILD revolutionizes how AI engineers in manufacturing and predictive maintenance build models by shifting from data-heavy workflows to a system that leverages human expertise. This approach enables AI engineers to capture and apply expert knowledge directly into models, reducing the reliance on vast amounts of data.

As the Founding Product Designer, I played a key role in shaping this product by leading user research, building a scalable design system, and establishing the design process that empowers engineers to build models faster and more accurately.

The Problem: Data-Heavy and Time-Consuming Workflows

Traditionally, AI engineers rely heavily on large datasets to build predictive models. However, many industrial environments—like predictive maintenance in manufacturing—struggle with limited, fragmented, or noisy data. AI engineers like Phuoc Bui and Zhang Yu-san faced constant delays caused by:

Insufficient Data: Engineers lacked enough high-quality data to train their AI models effectively, leading to inaccurate predictions.
Manual Data Preprocessing: The time-consuming process of cleaning and organizing incomplete data delayed model building.
Disconnection from Domain Experts: Translating domain knowledge from experts into AI models was inefficient and often involved back-and-forth communication, slowing down the process.

Before meeting with the client's AI engineer, Zhang Yu-san, I interviewed two internal AI engineers, Phuoc and Roshan, at Aitomatic. This gave me crucial insights into the daily frustrations of AI engineers and how they interact with data, domain experts, and model-building tools.

These interviews revealed several pain points: time-consuming data cleaning, difficulty incorporating expert knowledge into models, and the challenge of adjusting models for external factors like seasonal changes.

My interview with Zhang Yu further confirmed these challenges:

Disconnection from Domain Experts: Zhang rarely interacted with the subject matter expert after the project kicked off, which created delays when incorporating domain-specific knowledge.
Manual Data Preprocessing: Zhang spent hours cleaning and organizing inconsistent IoT data, taking time away from building and refining models.
Model Complexity and Environmental Adjustments: Due to seasonal variations in equipment behavior, Zhang had to adjust his models manually to ensure they remained accurate, which added another layer of complexity to his work.

These insights were invaluable in shaping the direction of K1st BUILD.

My interview with Zhang Yu further confirmed these challenges:

Disconnection from Domain Experts: Zhang rarely interacted with the subject matter expert after the project kicked off, which created delays when incorporating domain-specific knowledge.
Manual Data Preprocessing: Zhang spent hours cleaning and organizing inconsistent IoT data, taking time away from building and refining models.
Model Complexity and Environmental Adjustments: Due to seasonal variations in equipment behavior, Zhang had to adjust his models manually to ensure they remained accurate, which added another layer of complexity to his work.

These insights were invaluable in shaping the direction of K1st BUILD.

Design Approach: Leveraging Human Expertise Over Data

The vision for K1st BUILD was to shift the focus from data to human expertise, enabling engineers to capture and apply expert knowledge directly into AI models. I led the user research phase by interviewing AI engineers (Phuoc at Aitomatic and Zhang Yu-san at CNA) and domain experts to understand how knowledge, rather than data, could drive model-building processes.Key insights from the research included:

AI engineers needed a systematic way to capture and translate expert knowledge into models.
Reliance on large datasets slowed down the AI development process, and engineers required a way to build models with limited data but rich domain expertise.
Domain experts were often disconnected from the AI workflow, making it difficult for engineers to apply nuanced, expert knowledge to improve model accuracy.

K1st BUILD was designed to solve these critical workflow issues:

Streamlined Knowledge Sharing: K1st BUILD’s centralized knowledge library allowed Zhang to access structured domain knowledge from experts in real-time. This eliminated the need to wait for email responses or schedule meetings to retrieve vital insights, allowing Zhang to apply domain expertise as he developed models.
Automated Data Preprocessing: The tool automated the data cleaning process, turning raw IoT data into ready-to-use formats for Zhang. This saved him hours of manual work and ensured that the data was consistent, clean, and reliable, improving overall model performance.
Adaptive Model Building:
With K1st BUILD, Zhang could develop models that automatically adjusted to environmental factors, such as seasonal changes in equipment performance. This eliminated the need for Zhang to manually tune thresholds, allowing his models to maintain accuracy without constant intervention.
Cost-Efficient Model Monitoring:
K1st BUILD reduced cloud infrastructure costs by streamlining the model retraining process. Zhang could efficiently test and deploy models without straining his cloud resources, cutting down on both time and expenses.

K1st BUILD was designed to solve these critical workflow issues:

Streamlined Knowledge Sharing: K1st BUILD’s centralized knowledge library allowed Zhang to access structured domain knowledge from experts in real-time. This eliminated the need to wait for email responses or schedule meetings to retrieve vital insights, allowing Zhang to apply domain expertise as he developed models.
Automated Data Preprocessing: The tool automated the data cleaning process, turning raw IoT data into ready-to-use formats for Zhang. This saved him hours of manual work and ensured that the data was consistent, clean, and reliable, improving overall model performance.
Adaptive Model Building:
With K1st BUILD, Zhang could develop models that automatically adjusted to environmental factors, such as seasonal changes in equipment performance. This eliminated the need for Zhang to manually tune thresholds, allowing his models to maintain accuracy without constant intervention.
Cost-Efficient Model Monitoring:
K1st BUILD reduced cloud infrastructure costs by streamlining the model retraining process. Zhang could efficiently test and deploy models without straining his cloud resources, cutting down on both time and expenses.

How K1st BUILD Solved These Challenges

K1st BUILD allowed AI engineers to build models by relying on human expertise, rather than massive amounts of data. This was achieved through the following key features:

Expert Knowledge Capture: K1st BUILD allowed engineers to capture and apply expert knowledge directly into their models. This reduced the need for data-heavy model building and empowered engineers to leverage domain expertise more efficiently.
Knowledge-to-Model Translation: K1st BUILD provided a streamlined process for translating domain knowledge into AI models, using a symbolic approach. From the raw knowledge in text, AI engineers could have expert rules, heuristics to build AI models.
Adaptive Models: K1st BUILD enabled engineers to update their AI models with new knowledge inputs from experts, making the system adaptable and reducing the need for manual data searching.

K1st BUILD allowed AI engineers to build models by relying on human expertise, rather than massive amounts of data. This was achieved through the following key features:

Expert Knowledge Capture: K1st BUILD integrated tools that allowed engineers to capture, structure, and apply expert knowledge directly into their models. This reduced the need for data-heavy model building and empowered engineers to leverage domain expertise more efficiently.
Knowledge-to-Model Translation: The platform provided a streamlined process for translating domain knowledge into AI models, using a symbolic approach. Engineers could input expert rules, heuristics, and knowledge sets into the system, which then applied these insights to build smarter models without relying on large datasets.
Adaptive Models: K1st BUILD enabled engineers to create AI models that could be easily updated with new knowledge inputs from experts, making the system adaptable and reducing the need for constant manual adjustments.

I co-worked on the UI with front-end engineers, ensuring a smooth handoff with a clear design checklist. Within three months of joining Aitomatic, I created a design system for K1st BUILD, incorporating user feedback and insights.

K1st BUILD allowed AI engineers to build models by relying on human expertise, rather than massive amounts of data. This was achieved through the following key features:

Expert Knowledge Capture: K1st BUILD integrated tools that allowed engineers to capture, structure, and apply expert knowledge directly into their models. This reduced the need for data-heavy model building and empowered engineers to leverage domain expertise more efficiently.
Knowledge-to-Model Translation: The platform provided a streamlined process for translating domain knowledge into AI models, using a symbolic approach. Engineers could input expert rules, heuristics, and knowledge sets into the system, which then applied these insights to build smarter models without relying on large datasets.
Adaptive Models: K1st BUILD enabled engineers to create AI models that could be easily updated with new knowledge inputs from experts, making the system adaptable and reducing the need for constant manual adjustments.

Key Workflow Improvements: Before and After K1st BUILD

Before K1st BUILD, AI engineers like Phuoc and Zhang Yu were bogged down by incomplete datasets and manual data preparation tasks. They had to repeatedly rely on back-and-forth communication with domain experts to extract knowledge and apply it to their models.

After K1st BUILD, these AI engineers can apply expert knowledge, bypassing the need for large datasets. The knowledge system provided real-time access to expert insights, and applied these insights to model development, eliminating operational delays.

Prototype and Feedback

After presenting the early prototype of K1st BUILD to Zhang Yu and his team, I gathered detailed feedback that shaped the final design.

Here’s what we learned:

Zhang Yu found the tool easy to use for tuning existing models and integrating new data streams. The simple interface allowed him to implement changes without needing extensive support.
Guan-san, the domain expert, appreciated how quickly and easily he could share his knowledge with the engineering team. This integration made it possible for Zhang to work more independently.

Impact

The introduction of K1st BUILD had a significant impact on how AI engineers built models, as it fundamentally changed the way they worked by prioritizing expert knowledge over data:

With expert knowledge capture tools, Zhang Yu was able to build models 30% faster, as he no longer needed to wait for large datasets or spend hours cleaning data.
Models built with expert knowledge were 15% more accurate, as they incorporated domain-specific rules and insights that data alone couldn’t provide. This led to more reliable predictions and better outcomes in predictive maintenance.
K1st BUILD created a collaborative space where engineers could seamlessly integrate expert knowledge into their models, strengthening the relationship between them, resulting in higher-quality models and more efficient workflows.

Reflection

My role in K1st BUILD was to design a tool that not only simplified the AI engineer’s workflow but also fundamentally shifted how models are built—from a data-centric to a knowledge-centric approach. I ensured that K1st BUILD would position AI engineers to build more accurate models without the need for extensive datasets.

This project reinforced my belief in the power of human-centered design, as K1st BUILD’s emphasis on human expertise proved that thoughtful design can significantly improve the productivity and experience, even in highly technical environments.