I'm an AI Specialist at Kickstart AI. My role is quite versatile, and I switch between data science and machine learning engineering, depending on the project.
For example, in our recent project with Ampere, I took on more of a data science and statistician role. We were working on predicting parcel volumes, which required careful statistical modelling to ensure accurate predictions while building in safety margins.
For our food bank chatbot project, I was more involved in the machine learning engineering aspect. We developed a chatbot to help people determine their eligibility for food bank services. And we focused on creating an empathetic tone and overcoming language and cultural barriers.
What I love about these challenges is that we're not just creating models theoretically or in isolation. We're thinking about how they translate into real-world applications and improvements.
I studied computer science and econometrics, which led me to machine learning. I've always been fascinated by the beauty of mathematics. Especially how statistical methods can be used to reveal incredible insights.
During my bachelor's thesis, I used neural networks in a financial context. I then started looking at language models. What I love about AI is that it applies the mathematical methods I learned—just on a much larger scale.
Back in 2019, before ChatGPT, there were already language models that would predict next tokens. I remember being super impressed simply by their ability to predict the next word in a sequence. How far we've come in such a short time!
After university, I started here at Kickstart as a machine learning engineer.
I believe AI is the biggest revolution since the Industrial Revolution. What motivates me is the ability to apply these cutting-edge technologies to real world applications. And further, to get a nuanced view of their capabilities and limitations. Working hands-on with AI, you gain a much more complete understanding of it than just following the latest news.
I also really enjoy the engineering side - working towards a tangible product or project that you can actually showcase and interact with. That's something I was missing in my academic work, and it's very satisfying in my current role at Kickstart AI.
One project I'm particularly excited about was one which involved analysing the news using topic modelling. We used a dataset of 100,000 articles from Dutch news sources over the last 10 years. We applied BERT-topic, a method that uses the knowledge embedded in large language models to position articles in an abstract space and cluster them by theme.
This allowed us to discover trends and patterns in news coverage over time. For example, we could see spikes in certain topics like healthcare becoming more prominent around election times. It's fascinating to see these patterns emerge from the data.
You can read more about the project and workshop about it that I ran here.
I mostly use Claude for chatting and coding assistance. It's particularly good at providing complete code snippets for things like web programming (I find it much better than ChatGPT). It really accelerates experimentation and helps to quickly develop minimum viable products.
I also use Cursor, an AI-integrated IDE that allows for seamless interaction with code. You can chat about your code, suggest edits and get explanations. I find it makes me much more efficient at coding.
I often use these AI tools to quickly prototype ideas or build simple websites to showcase concepts.
I'm a bit addicted to Hacker News - I check it 5-10 times a day! I find the quality of discussion there to be much higher than on other social networks. It's a great place to get interesting perspectives on the latest AI developments.
I also follow some key AI figures on Twitter, like Andrej Karpathy. But Hacker News is my main source for staying on top of AI news and trends.
I'm really interested in the advances in AI for medicine. The work that companies like DeepMind are doing in combining AI with biotechnology, like AlphaFold for protein structure prediction, is fascinating. These techniques could potentially accelerate drug discovery and lead to breakthroughs in treating chronic illnesses.
I'm also excited about the further integration of multimodal models. As models like GPT-4 get better at reasoning across text, images, and audio, we'll see some really interesting applications. Imagine a model that can reason with diagrams, or provide language learning feedback based on your pronunciation. The possibilities are endless.
Even in an applied AI role, it's valuable to know how these models work, at least a medium level of detail. Understanding concepts like the attention mechanism, tokenization, or positional encoding helps you work better with these models. And also to understand their limitations.
For example, understanding how tokenization works helps you understand why GPT-4, a super strong model, is not able to determine that there are three 'r's in ‘strawberry’. If you understand how tokenization works, and you know that actually what the model sees is not letters, but rather parts of words. Then it all makes sense.
💡 Want to learn more about Kickstart AI and our mission to accelerate AI adoption in the Netherlands? Check out our website, read more about Cascha’s project analysing the news, or connect with him on LinkedIn.
