.jpg)
TechDays 2024 was held over two days, the 29-30th May, at Thuishaven in Amsterdam. We brought leading experts to the stage, and explored exciting AI developments. The whole event was full of practical insights, engaging workshops, networking, and cross-company collaboration. The Thuishaven venue added to the special atmosphere of the event.
There were many highlights at TechDays, here are some of them:
Speakers: Martynas Jašinskas & Jokūbas Krasauskas, Vinted
Martynas and Jokūbas flew in from Lithuania especially for this keynote, and kicked off TechDays 2024 in style, with a riveting session about how Vinted combats fraud.
Vinted, a popular second hand marketplace, faces a constant battle against fraudulent activity. With its massive user base, of over 30 million daily active users, it has become interesting for malicious actors, and is often targeted by organised crime groups.
Martynas and Jokūbas explained how their main goal is to protect the Vinted platform and its users from these threats. The continuous evolution of fraud requires constant improvement and adaptation. They highlighted the sophisticated machine learning models and detection tools they use.
Vibe: The room buzzed with excitement and curiosity. Attendees were super engaged, nodding along with the deep insights that were being shared. (We later heard from Vinted that the speakers shared things that shouldn’t be shared publicly, so attendees got a real insider’s look / treat!)
What Stood Out: The challenge of real-time adaptation of machine learning models to adapt to evolving fraud patterns. And the practical applications and effectiveness of Vinted’s data-driven solutions.
Speakers: Simon Grest & Ruben van de Geer, Albert Heijn
Simon and Ruben demonstrated how they use AI for hourly demand forecasting, and production planning, for Albert Heijn's in-store bakeries. Their algorithms ensure product availability (i.e. enough bread), while significantly reducing food waste. Attendees seemed to appreciate the behind-the-scenes look at AH’s data-driven bakery operations.
What Stood Out: The delicate balance between forecast accuracy and operational efficiency.
Speaker: Branka Milivojevic, NS
Branka shared how optimization models have helped NS save energy at train stations, balancing their own business rules, efficiency, and costs. She also presented NS’s sustainable three-year plan. Attendees seemed to value the data-driven ways that were shared to solve environmental issues.
What Stood Out: The use of optimization models in real-world sustainability. How NS has managed to achieve a tangible balance between their various constraints.
Speaker: Cascha van Wanrooij, Kickstart AI
Cascha introduced how BERTopic can be used to analyse news trends, providing hands-on experience with topic modelling on a Dutch news dataset. The workshop also explored socio-economic indicators related to poverty. In this interactive and hands-on workshop, participants worked actively on their laptops.
What Stood Out: The potential of topic modelling in news analysis and its creative applications to broader social issues. Get more detail in Cascha's blog post on this topic.
Speaker: Max Baak, ING
Max introduced the Entity Matching Model (EMM) Python package for handling large datasets, providing detailed performance numbers, and an engaging demo.
What Stood Out: The scalability and efficiency of the EMM, and its practical applications in large-scale data handling.
Speaker: Evangelia Toutou, Kickstart AI
Evangelia explored the complexity of translating sign language, due to its unique linguistic structures, and discussed the fine-tuning of multi-modal transformers for this task.
What Stood Out: The intersection of NLP and social good, and the innovative use of multi-modal transformers in an unfamiliar field.
Speaker: Koen de Raad, Eyedle
Koen discussed using real-time computer vision for recording and analysing field hockey matches. He addressed challenges in setting up the hardware, in synchronising it, and in detecting both players and the ball.
What Stood Out: The practical implications of real-time computer vision, and its potential to revolutionise sports analytics.
Speaker: Turan Bulmus, Google
Day 2 of TechDays kicked-off with a keynote about the ever-evolving GenAI landscape. With over a decade of experience in developing AI technologies, Turan had a strong background to draw upon for this session.
It was an insightful and forward-thinking session about navigating the journey from experimentation to production with Generative AI, specifically within large GCP environments. It covered key challenges in production (such as data quality and model bias), and practical strategies for building and deploying GenAI-driven applications, Turan also talked about how to bridge the gap between MLOps and LLM Ops, and shared real-world insights and lessons learned from successful GenAI implementations. He emphasised the need for continuous retraining, model versioning, and the inclusion of human feedback in fine-tuning processes.
What Stood Out: The shift towards application development in AI, and the sophisticated use of LLMs to verify other LLMs.
Speaker: Omendra Manhar, Kickstart AI
This workshop demonstrated that Retrieval Augmented Generation (RAG) is not as mythical as it sounds. Participants created a chatbot that uses RAG to answer questions about the Belastingdienst.
This was a collaborative and practical session, with participants engaging in hands-on activities on their laptops. Key takeaways included how easy it is now to create an initial application and the importance of pre-processing data.
What Stood Out: The workshop highlighted the truth of the phrase "garbage in, garbage out", and stressed the need for good data preprocessing in RAG apps.
Speaker: Nastasha Govender-Ropert, ING
Nastasha's session shed light on the consequences of missing data for women across various domains, leading to gender inequality. She discussed historical biases, economic disparities, and the unconscious bias in studies and data interpretation.
What Stood Out: The session's call to "change the system, not the women" emphasised the need for systemic changes to address gender bias.
Speaker: Rutger Brouwer & Sander Boot, KLM
This session showcased how ML models can enhance decision-making processes to prevent aircraft delays. The key takeaway was the challenge of making models ready and accepted by operations.
What Stood Out: The importance of domain knowledge in improving model performance and operational acceptance.
Speaker: Madhusudhanan Srinivasan, bol
Madhusudhanan discussed how to target an audience with relevant display ads using propensity models. The session included a detailed explanation of predicting customer behaviour and addressing audience questions.
What Stood Out: The practical application of propensity models in predicting customer behaviour for targeted advertising.
Speaker: Maarten Grootendorst, BERTopic Creator
BERTopic is a topic modelling technique developed by Maarten. It uses transformers and c-TF-IDF to make dense clusters. These clusters allow for easy-to-understand topics and keep topic descriptions full of important words. Maarten shared his personal journey of overcoming challenges and optimising for efficiency due to limited resources, such as chronic pain, and limited GPU power. His inspirational talk focused on making limited resources work to your advantage.
Vibe: Motivational and personal, leaving attendees inspired.
What Stood Out: The emphasis is on problem-solving and stakeholder thinking before coding. The idea is that having limited resources can lead to greater efficiency and creativity.
Read more about BERTopic in Maarten’s Medium post.
TechDays was a big success. It highlighted the real-world applications of AI and data science, and inspired attendees to push the boundaries. Thanks to everyone who attended, presented, and contributed. Your energy and eagerness to learn made TechDays 2024 special.
And we can’t wait for the next one!
> See all our upcoming community events here.
