Introduction
Shell is growing exponentially in the E-Mobility space! As you may already know, Shell has been committed to expand its electric vehicle (EV) charging network from 60,000 now to 500,000 chargers in 2025 across the world. This is exactly why we want to motivate you to come work with us on this year’s challenge that will help Shell develop more intelligent and efficient way of operating our EV charging sites.
‍
Challenge
For this case, we zoom to an EV charging site where a 300 kW/360 kWh stationary battery is deployed to support EV charging operations. This site has two DC chargers, and the stationary battery is an integral part of the local charging system. Inclusion of batteries in overall workflow introduces unique challenges as well as opportunities. While there is a cost associated with every battery charged and discharge operation, it can also be used for temporal arbitrage. Your challenge is then to develop an AI-driven solution to maximize both profit and battery lifespan depending on several factors, such as charging demand, electricity price and battery characteristics.
‍
What we ask of you
• Full commitment to the competition and minimal 1 day per week (on the agreed day) to our site at Shell’s Energy Transition Campus Amsterdam.
• Students in STEM related field with experience in ML/AI and good programming and communication skills.
• Compensation provided and travel expenses are reimbursed if you are not in possession of student OV card.

Introduction
Steel surface quality is one of the most important product properties to our customers (particularly our automotive customers). Hence, there are camera inspection systems placed in the production lines of Tata Steel’s that continuously scan the steel surface for defects. These systems are used to detect and classify surface defects and feed the systems that determine whether the steel meets the quality standards requested by our customers. In some cases, the coil needs to be degraded – in other cases part of the coil needs to be cut out.
Based on experience we know that the defect classification routines built-in the software provided with the camera hardware is not very accurate. The last couple of years, we have successfully applied deep-learning methods to improve this classification. However, we face a number of bottlenecks to truly scale this up:
1. Labelling: we require typically >10k images which takes a significant amount of time from our surface quality experts.
2. Accuracy: despite having labelled >10k images, the model precision typically reaches a maximum plateau around 95%.
3. Parsimony: we currently use a two-step modelling framework (CNN and subsequently a Random Forest) to combine defect image data with tabular data of this image (position on the strip). This is inefficient when it comes to model maintenance, and performance-wise probably suboptimal.
4. Re-useability: camera hardware, settings and steel appearance hamper the re-usability of the developed deep learning models. This greatly increases the required labelling efforts.
On all these aspects, we have ideas to improve our current way of working. For instance, we are very curious whether object detection instead of image classification can reduce the labelling burden. Also the use of unsupervised learning techniques in the labelling process might drastically reduce the labelling effort. The accuracy of our models might be increased by better accommodating for the considerable class imbalance. To reduce the number of models to maintain, we think about a solution in which auxiliary variables are somehow included in the convolutional neural network. On the re-usability, we have been hypothesizing that Generative Adversarial Network (specifically cycle-GANs) might be a way to re-use labelled images measured under different conditions.

Challenge
Our challenge to you: can you develop a Proof-of-concept on one or more of these topics? We deliberately present a wide range of topics because we want to explore together with the students on which topics they think there is the most to gain. And there is certainly not a lack of data on which to test and trial new methods: we have a cloud environment packed with >100k images at your disposal! By helping us improve steel defect monitoring, you can directly contribute to Tata Steel’s overall quality performance!

Your summer of AI at Tata Steel
During the summer, you’ll be coached by Tata Steel advanced analytics coaches and gain experience in our agile way of working. You'll have contact with process matter experts to validate preliminary results.
At the end of the summer, you have learned how to create a proof of concept and work with large scale data systems, such that at the end of the summer, we can start implementing this system.

What we ask of you
• 4 days per week (non-weekend) of which min. 1 day per week on our site at the Tata Steel IJmuiden.
• Preferred master student in STEM related field with experience with ML/AI.
• Compensation provided and travel expenses are reimbursed if you are not in possession of student OV card.‍

Introduction 
As a part of the FD Mediagroep, Company.info holds the position of the market leader in providing Dutch business information and insights. Our clients rely on us for qualitative, up-to-date, and in-depth insights into the latest company news. We cater to the needs of every business professional who wishes to establish robust business relationships and boost their sales. Our data is accessible via various channels, including an online portal, APIs for system integration, online dashboards for analysis, and data delivery. By providing these solutions, we aim to help professionals achieve more success in their business endeavors.

‍Challenge 
We want to link companies' websites to legal entities from Kamer van Koophandel (KVK) and to the buildings in Basisregistratie Adressen en Gebouwen (BAG). An important piece of information we use for this are addresses. We found that currently there’s no Deep Learning model that is good at extracting Dutch addresses from text. We want to train a model that extracts Dutch addresses well. At Company.info we have data of millions of Dutch companies and their websites, including large amounts of hand-labelled data. For this case, we have data consisting of pairs of websites and addresses that we extracted using regular expressions. Training the first effective Dutch address extractor would already be quite a feat, but if time allows there are different options for expanding upon that, depending on the interests and creativity of the team.
‍
‍Your summer of AI at Company.info 
You will work closely with our data science team of four and have dedicated supervisors. You will gain practical experience in applying state-of-the-art Deep Learning / NLP techniques for solving real-world problems as a Scrum team. You will also enjoy the friendly atmosphere of Company.info and are welcome to join our weekly drinks.

‍What we ask of you
• A solid foundation of Python knowledge
• Experience with Deep Learning, NLP and git is a plus
• Enthusiasm for teamwork
• We expect you to be at our Amsterdam office 2-3 days a week in order to meet your team face-to-face

What we offer
• Market rate internship compensation
• Travel expenses are covered
• A laptop to work on
• An office that’s easy to reach from Amstel Station
• Supportive and fun work environment in a mid-sized company
• Enthusiastic and helpful colleagues around you

Your task is to uncover interesting patterns to explain or predict these food insecurity phases in regions in the Horn of Africa. We will provide you with datasets from classical sources, including monthly climate and weather data, food prices from the WFP, and conflict data from an academic database. Additionally, we will supply standardized data from common news sources. Prior research has shown that NLP methods applied to news sources can predict food insecurity in advance [BSF23], and we are curious to see if you can expand upon this. Lastly, we encourage you to develop new and innovative ideas to track or predict food insecurity using public data we haven’t considered before, such as data from satellite images.

What We Ask of You
• A background in data science/machine learning or other technical domains (e.g., master students of AI, Econometrics, Data Science, Computer Science, etc.), or strong data science skills. Programming knowledge is required, such as working with Python libraries like Pandas and Sklearn or equivalents in other languages.
• Some background knowledge in NLP (classical methods or Deep Learning).
• An interest in a project that considers context, rather than being solely technical. For example, you might want to learn about the recent history of the Horn of Africa to better understand the situation.
• Experience with applying data science methods on real-world data. For instance, working with sub-optimal data containing missing values and requiring data processing. As a scientific lab, we are not seeking a perfect solution, but rather interesting insights!

What We Offer
• A fun, startup-like environment. You will work at our brand new Delft Kickstart AI office a minimum of one day per week, although you are welcome everyday of the week. We expect you to be available four days per week to work on this project. Our office is complete with a playstation, table tennis tables, free coffee and tea and of course, weekly borrels!
• The opportunity to become part of Kickstart AI’s data science community, with specific regular engagement with our four partner companies: ING, Ahold Delhaize, NS and KLM. You will receive invitations to attend Kickstart AI’s exclusive events, speak with Heads of AI from our partner companies and engage regularly with their data scientists and engineers.
• Work with both corporate companies, and universities, on a project that is focused on applied AI, for good. In joining our efforts, you will not only expand your skill set and knowledge, but also contribute to a meaningful mission!
‍
References 
[BSF23] Ananth Balashankar, Lakshminarayanan Subramanian, and Samuel P. Fraiberger. Predicting food crises using news streams. Science Advances, 9(9):eabm3449, 2023.

Problem statement
Our sales team regularly meets with customers and records the details of each visit in the CRM system. We want to analyse these reports to identify common customer needs and preferences, so that we can better tailor our sales approach and improve our sales effectiveness.

‍Data collection and pre-processing
We collect customer visit reports from the CRM system and pre-process the data to remove irrelevant information such as timestamps, sales rep names, and other identifying information. We also clean the text data by removing punctuation, stop words, and performing stemming or lemmatization as necessary.
‍
‍Data analysis
We use NLP techniques to analyse the text data and extract key insights. Here are some possible analysis techniques:
• Topic modelling: We can use topic modelling techniques such as Latent Dirichlet Allocation (LDA) to identify common themes or topics that appear in the customer visit reports. This can help us understand what customers are most interested in and what their top pain points are.
• Sentiment analysis: We can use sentiment analysis techniques to identify whether customers are expressing positive, negative, or neutral sentiment in their visit reports. This can help us understand how customers feel about our products or services and what we can do to improve their experience.
• Named entity recognition: We can use named entity recognition techniques to identify specific products, features, or services that customers are mentioning in their visit reports. This can help us identify which products or features are most important to customers and what we should focus on in our sales pitch.

‍Insights and action
Once we have analysed the customer visit reports, we can use the insights to improve our sales effectiveness. Here are some possible actions we could take:
• Tailor our sales pitch: Based on the common themes and topics identified through topic modelling, we can tailor our sales pitch to better address customer needs and pain points. For example, if customers frequently mention a need for better customer support, we can emphasize our strong customer service in our sales pitch.
• Improve our products or services: Based on the sentiment analysis results, we can identify areas where customers are expressing negative sentiment and take steps to improve those aspects of our products or services. For example, if customers are frequently complaining about long wait times on the phone, we can improve our call centre processes to reduce wait times.
• Focus our sales efforts: Based on the named entity recognition results, we can identify which products or features are most important to customers and focus our sales efforts on those areas. For example, if customers are frequently mentioning a particular product feature, we can make sure our sales reps are trained to highlight that feature in their sales pitch.