Harry Wik

Neo4j is a graph database company that helps organisations store and analyse connected data for applications such as recommendations and fraud detection. It also provides graph analytics and machine learning tools, alongside enterprise support for customers running its database in production.

Training machine learning models on large graphs can require substantial memory and data transfer, which limits the size of datasets teams can work with. Neo4j's Graph Data Science team wanted to connect its database with PyTorch Geometric so models could use graph data without loading the entire graph into memory. The project focused on making training and inference practical for larger datasets while reducing resource requirements.

I joined as a data scientist, built and compared several pipelines for training and inference. I measured batches processed per unit of time and timed individual stages, including sampling, to identify bottlenecks. I improved data representations, caching and database procedures to reduce data transfer and memory requirements. A colleague and I also built a plugin that ran graph convolutional network predictions directly within Neo4j without an external machine learning library. The team increased pipeline throughput 3-4x, with tests covering graphs of up to 110 million nodes and 1.6 billion relationships. The work enabled teams to train models on larger graphs with less memory and run predictions directly in Neo4j without a separate inference service.

Neo4j is a graph database company that helps organisations store and analyse connected data for applications such as recommendations and fraud detection. It also provides graph analytics and machine learning tools, alongside enterprise support for customers running its database in production.

When security vulnerabilities affected their installations, customers needed to know which Neo4j versions would resolve them without introducing compatibility problems. An upgrade could also require changes to drivers, plugins and storage formats, making guidance dependent on several related products. The support team wanted a tool that could connect this information and calculate compatible upgrade paths, reducing the manual work involved in recurring customer inquiries.

I joined the support enablement team as a data engineer. I designed a graph model linking product versions, software dependencies and security vulnerabilities so the application could calculate compatible upgrade paths. I built the Java backend with Spring Boot, the user interface and an OpenAI integration that summarised important upgrade changes. The application brought compatibility checks, upgrade guidance and security notes into one tool. NeoDAVE remains an ongoing project at Neo4j.

The client was a research collaboration focused on critical care medicine.

Predicting intensive care mortality risk can support clinical decision-making, but clinicians also need to understand the reasons behind a prediction. The research team wanted to evaluate whether interpretable machine learning models could provide useful predictive performance while making each risk estimate auditable. The project examined how individual patient characteristics and their interactions could explain the model's predictions.

My team and I evaluated interpretable machine learning models on the MIMIC-III intensive care dataset. I compared implementations and assessed their predictive performance and interpretability. The models accounted for individual patient characteristics and interactions between pairs of characteristics. The system automatically generated log-odds diagrams that made each risk estimate auditable by showing how these factors contributed to the prediction.

Hitachi Energy is a global technology company that develops products and systems for electricity generation, transmission and distribution. Its work includes power grids, transformers and high-voltage equipment, helping utilities and industry supply reliable electricity and integrate renewable energy.

The requirements engineering team worked with internal documents and relevant ISO and IEEE standards to assess compliance. It wanted to explore whether generative AI could reduce the manual effort involved in finding and interpreting information across these sources. Because some documents contained sensitive information, the prototype had to run entirely on-site.

I built and evaluated a retrieval-augmented generation prototype using a Mistral model. The system used internal documents and applicable standards as source material for its answers. I made the design modular so the team could replace the embedding and language models as better options became available. The prototype helped the team understand the technology's capabilities and limitations and provided a basis for evaluating newer models.

Careerbia was a Swedish recruitment startup that aimed to make hiring simpler for job seekers and small-business employers. Its platform connected candidates with employers through reusable profiles and direct communication, reducing the need for repeated applications.

Applying for jobs often requires candidates to enter the same information repeatedly, while employers need a way to identify suitable applicants and start conversations. Careerbia aimed to simplify recruitment through reusable applicant profiles, job matching and direct communication. The platform needed to support job discovery and applications for candidates, alongside vacancy publishing and applicant search for employers.

I co-founded Careerbia and designed the initial data model for profiles, job matching and chat. I developed backend services on AWS, helped rebuild the REST API and contributed to the mobile applications and employer dashboard. I led pre-release testing with external users and coordinated fixes based on their feedback. The team collected over 200 contact sign-ups for a launch reminder, but slow employer recruitment prevented the platform from gaining commercial traction.