During ICT&Logistics, Sebastian Piest, Assistant Professor at the University of Twente, took us on an AI expedition with the aim of kick-starting an AI project. During his keynote, he introduced the Community of Practice for AI in Logistics, the toolbox developed for AI projects and several practical examples. In this article, we share the key take-aways.
The Community of Practice for AI in Logistics is an initiative from the University of Twente, Breda University of Applied Sciences and Deltago. In 4 workshops, participants develop their own idea into a design, prototype and implementation (plan). More than 30 companies have now participated in the AI expedition. See ai-expedition.nl
How do you kick-start a data science or AI project in Logistics?
Within the AI Expedition, together with participating companies, a practical and effective method was developed that helps you initiate a data science or AI project in four steps.
The first step is to organise a half-day design workshop to work out use cases together with a design canvas. Then, in the second step, you work out the design canvas in a design sprint into a mock-up and/or prototype. This allows you to visualise the end product and verify the idea with experts and end users. Depending on existing knowledge and skills, you can complete a design sprint in 3-5 days. Based on the results of the design sprint, you can make a good go/no go decision to start an AI project. Based on a business case, you work out a project plan using existing project methodologies such as CRISP-ML(Q).
Some AI projects can be realised in weeks, but more complex projects can take months or even years. Besides technology, implementation within the organisation is the most important step towards actual use. The fourth step focuses on AI implementation, where the AI application is put into use and monitored. From monitoring, a redesign process can be started.
How to get from AI prototype to implementation?
In many cases, developing an AI prototype can be done in isolation, with no/few integrations with existing systems required and minimal impact on processes. Making an AI application operational requires embedding it within existing or new processes and the IT landscape. There is often a gap between the prototype requirements and the requirements of an operational AI application, necessitating a redesign or alternative design. Therefore, it is important to develop a solution architecture that visualises the key components of the AI application and integrations with systems within the existing IT landscape.
Equally important is to carefully design users' interaction with the AI application and adapt processes accordingly or design new ones. Since data are not static and change over time, the performance of AI models will need to be monitored and periodically retrained. So it is not a traditional implementation process, but more of a path of continuous improvement.
What can we learn from previous AI projects?
From the Community of Practice for AI, we see several ideas for AI applications popping up. These include predicting demand and volumes with machine learning, processing data and documents with large language models, visual inspection of pallets in warehouses with computer vision to promote safety, optimising tank decisions with reinforcement learning, master data checking with rule-based AI, and many other examples. These use cases serve as inspiration for new participants and, in some cases, result in reusable code. By starting an AI expedition together, we see that participants help each other purposefully advance in the search for AI models and a lot of cross-fertilisation occurs, especially when students and IT partners are involved and AI models are shared between them.
Developing an AI prototype generally succeeds well, but actually implementing it, putting it into use and developing it further does not work in many cases, or only to a limited extent. Furthermore, we see that AI is not always the best way to solve a problem and the performance of AI models is insufficient for user acceptance. Sometimes statistical models or operations research techniques are better, faster and easier to implement. Therefore, we have broadened the initial focus on AI projects to include data science as well as AI projects.
How do we increase the success rate of an AI project?
AI projects usually start with a lot of energy and enthusiasm, but often also with unrealistic expectations, uncertainty as to whether AI models are suitable for the task/problem and uncertainty as to how the AI application will be embedded in the IT landscape and used operationally. To create realistic expectations, it is important to quickly concretise and verify ideas for new AI applications with experts and end users. Creating a solution architecture helps here to create clarity and give direction to the development path.
Furthermore, there are important preconditions that need to be fulfilled, including the availability of sufficient and high-quality data, a team with knowledge of processes, IT and AI, a sponsor and a concrete objective and budget. Performing a design sprint helps to determine the scope, size and complexity, and also to identify the risks and preconditions for an AI project. This does not guarantee success. As more and more case studies are published, we can move towards best practices and reusable AI applications.
Which generic algorithms can add value when starting an AI project?
At the start of an AI project, it is often unclear which algorithm is the best fit for the issue and whether subsequently the necessary data is available. It is valuable to be able to quickly test and compare different AI models at the start of an AI project. There are several open source tools and libraries available with algorithms for (un-)supervised machine learning (e.g. Skikit-Learn, XGBoost), deep learning (e.g. TensorFlow, PyTorch) and large language models (e.g. LangChain).
There are also several commercial AI products and AI platforms that can be used to develop and implement algorithms. Don't stare blindly at the algorithm, because more is needed to develop an AI application. Besides algorithms, it is wise to look closely at deployment and integration options, available documentation, training and support via forums or implementation partners.
Can you elaborate on the toolbox that can help to start your own AI expedition?
With the toolbox for AI projects, you can go through the different steps quickly and in a focused way. An AI expedition starts with an intake, in which the goals, possible use cases, team, existing knowledge and tools are inventoried. You then use the design canvas to work out the idea behind the use case. You then create a visualisation using the solution concept / architecture template.
For mock-ups and prototyping, we use the 18 Human-AI eXperience (HAX) Guidelines, Workbook and Playbook and tools such as Miro and Figma. To collect user feedback, we have developed scorecards with commonly used KPIs. To manage AI projects, we use CRISP-ML(Q) as a project methodology. There are Python Notebooks that follow this methodology and provide standard (open source) algorithms. Alternatively, we increasingly see PowerBI being used as a starting point, which is nowadays equipped with Machine Learning / AI features and can be extended with Python scripts .
Additionally, we use tools such as Jira and GitHub to support software development. Finally, we are increasingly deploying AI-assisted development to speed up the process. This includes support for brainstorming, working out use cases, generating a project plan and writing and checking code. AI tools such as ChatGPT thus increasingly help you kick-start AI projects.
Author: Sebastian Piest
