During the 5th and 6th semesters of my computer science studies at the Hasso Plattner Institute, I had the opportunity to work on the GorillaWatch project. Together with a team of seven students, and in collaboration with Zoologischer Garten Berlin AG, Conservation X Labs, and SPAC (Sabine Plattner African Charities), we developed a machine learning system for the re-identification of western lowland gorillas.

The Goal:

Western Lowland Gorillas are a critically endangered species living in the Odzala National Park in the Republic of the Congo. To protect them, researchers need detailed insights into population sizes, individual movements, and social group behaviors. This information is vital for conservation efforts, particularly in preventing and managing disease outbreaks, such as the Ebola epidemic in the early 2000s, which wiped out approximately 70% of the gorilla population. 

Manually collecting this data is impossible due to the park's vast size and the large number of gorillas. Our goal is to create a re-identification system that can analyze video recordings from wildlife cameras and provide researchers with this critical information automatically.

Project Partners:

This ambitious project is supported by three key partners:

1. SPAC (Sabine Plattner African Charities
   Dr. Magdalena Bermejo and her husband German Bermejo, leading experts in gorilla research, have installed an extensive network of wildlife cameras in the Odzala National Park. They also manually labeled hundreds of videos, as there are very few publicly available datasets for gorillas.

2. Zoologischer Garten Berlin AG
   The Berlin Zoo is home to six gorillas. By setting up additional cameras in the zoo, we not only collected more training data but also created a controlled environment to validate and test our system.

3. Conservation X Labs
   Dante Wasmuth, an Oxford University alumnus and expert in computer vision, supported us with his expertise. He provided regular feedback on our progress and solutions during weekly meetings.

Technical Details:

Our system begins with a video input, which is divided into individual frames. Using a fine-tuned YOLOv8 model, we detect where gorillas appear in the video. Once detected, we extract body crops of the gorillas and filter them based on size, quality, and other criteria.

The filtered body crops are then processed by a video transformer model, specifically a fine-tuned DINOv2 model. At this stage, our system achieves an accuracy of 82%.

In addition to developing the re-identification pipeline, we are also building a web platform. This platform will enable researchers to access and share critical data, such as when and where specific gorillas were sighted, as well as insights into their social groups and movements.

My Bachelor Thesis:

My bachelor thesis focuses on the first step of the pipeline: extracting gorillas from the video frames. This involves two key tasks: detecting individual gorillas and tracking them consistently across frames. Tracking becomes particularly challenging in scenarios where two gorillas walk past each other or overlap, such as when a mother carries her child. These situations often result in ID switches, where the model either misidentifies a single gorilla as two different individuals or swaps the IDs of two gorillas.

In my thesis, I am investigating how the BoostTracking++ approach can reduce these ID inconsistencies and improve tracking accuracy in such complex scenarios. This work aims to enhance the overall reliability of the system, ensuring it can provide researchers with precise and actionable data for conservation efforts.

As part of my bachelor's studies at the Hasso Plattner Institute, I had the opportunity to develop a software product in a team of four during the seminar Trends and Concepts of Business Applications. Our client was Endress+Hauser, a global market leader in measurement and automation technology with several billion euros in annual revenue. The company operates in the B2B sector and serves industrial clients worldwide.

Our project focused on the final stage of the sales funnel — the post-sale customer experience. The goal was to use so-called O-Data (operational data), such as the customer's industry, company size, or country of origin, to predict what aspects of the product or service a customer might be dissatisfied with, such as customer support, delivery times, or product usability.

To enable this, we used historical customer feedback (X-Data) to train a machine learning model. This model could then forecast potential dissatisfaction points for new customers, allowing the company to proactively address those issues.

In addition to developing the software and writing a scientific paper documenting our approach, we also had the opportunity to present our product, called Anida, to a high-level audience of industry experts and company stakeholders.

During my final year of high school, I was recruited by SafeSpace founder Nour Idelbi to join the startup SafeSpace. In my role as CTO, I served as the bridge between the company’s vision and its technical implementation. Within just a few weeks, our team grew from two to five members, and we began developing the first prototype using React Native. 

The app’s idea is to help (young) women feel safer in uncomfortable situations, like walking home alone at night. Users can share their live well-being with pre-selected contacts through the app. In emergencies, they can also share their location and phone battery status or alert the police. The app also allows users to call others within the network, creating a community of people supporting each other to feel safer.

In addition to software development, I was responsible for coordinating the two other developers, as well as overseeing the design and strategic planning of the app’s features. Thanks to our strong presence on TikTok (at the time, we had already surpassed 10,000 followers), we had the opportunity to work closely with our target audience and make user-centered decisions regarding the app’s design and features.

In July 2022, we won the national startup competition STARTUP TEENS in the “Services & Platforms” category. Winning the highest-funded business plan competition for high school students not only earned us 10,000€ in prize money but also brought significant media attention.

Unfortunately, due to the start of my computer science studies in October 2022, I lacked the time resources to continue contributing to the project. However, I am deeply grateful for the opportunity to have been part of this initiative and to have made a significant impact.

As part of my Besondere Lernleistung during the final year of high school, I developed a app designed to help elementary school children learn to read. At its core, the app uses speech recognition to compare a text read aloud by a student with the reference text. Based on guidelines provided by the Ministry of Education — such as expected reading speed and allowed number of mistakes — the app calculates a score represented by a 1 to 5 star rating. If the score is high enough, the student advances to the next level.

By integrating gamification elements and personalization features, the app aims to optimize the user experience. From a technical perspective, it was developed as a native Android application using Java. All data is stored locally in an SQLite database to ensure data privacy.

In addition to receiving the highest possible grade (15/15 points) for the app and the accompanying academic thesis and oral defense, I was awarded multiple prizes in stratup and resaerch competitions such as Jugend forscht and Jugend gründet. This also led to various media opportunities to share the story behind the app.

The Fortune Cookie App was the first app I developed myself.
When the global COVID-19 pandemic hit in 2020, I used the time to teach myself app development through a Udemy course. As part of that journey, I created my very first app: a native Android application that allowed users to draw digital fortune cookie quotes.

By tapping on the fortune cookie—or shaking their phone—users could “open” the cookie and receive a randomly selected quote. The quote could be read or shared directly via WhatsApp. In addition to choosing from different categories, users were also able to create their own personalized messages.

Looking back, while it may not have been a technically complex project, I’m still proud of it as my first step into the world of software development.