Nina Žižakić

Nina Žižakić works on machine-learning applications ranging from image restoration to personalised running guidance and legal research. Her research with Aleksandra Pižurica developed compact image descriptors that made matching patches within a painting less memory-intensive. In her work and joint writing at In The Pocket, Nina has examined how to turn predictions and generated answers into applications people can test, question and correct.

Compact representations for image restoration

Nina earned bachelor’s and master’s degrees in computer science at the University of Novi Sad in 2014 and 2017, then completed her doctorate at Ghent University in 2022. Her doctoral research investigated compact image representations: encodings that retain enough information to find similar pictures or matching patches within an image. These comparisons support tasks such as image retrieval, tracking and filling missing areas in damaged images.

Working with Aleksandra Pižurica, Nina developed autoencoder-based image descriptors and tested them in virtual restoration of the Ghent Altarpiece. Their joint research found that variational autoencoders offered only marginal benchmark improvements over ordinary autoencoders while introducing more parameters to tune. They also modified the encoder architecture to reduce memory requirements when matching patches within a single image. In their experiment, the resulting method could process a whole painting panel where the comparison implementation exhausted the computer’s memory. Nina released the research code, including training and evaluation tools.

Models within running and legal workflows

By February 2023, Nina was working at In The Pocket and co-authoring guidance on developing machine-learning products. Her subsequent projects included predictive systems informed by specialist expertise and language-model applications built around existing documents.

For Start 2 Run, Nina worked with a team combining sports science, machine learning and product design. The running-guidance project analysed 7,500 running sessions and used an LSTM neural network to process sequences of speed and heart-rate measurements. The aim was to recommend a suitable pace for beginners without requiring each runner to undergo laboratory testing. Coaches’ methods informed the work, while the app translated predictions into running reports and actionable pace recommendations.

In her account of a legal-monitoring application, Nina addressed the daily work of checking Belgium’s Official Gazette for legislation relevant to clients. The team’s application collected new publications, used GPT to categorise and summarise legislation, identified references and amendments, and made the processed material searchable. Generation was one step in a workflow that also had to collect, organise and retrieve the underlying legal material.

Nina and Eleonora Drykina also jointly presented a payment-rulebook assistant built for an anonymous Belgian client, credited to In The Pocket. Its roughly 270 PDFs and 6,000 pages posed a retrieval problem: a question might require a table row, a whole section or several rulebooks. The team preserved section paths and chunk order so passages could be reconstructed, and converted tables into CSV while retaining the relationships between headings and values. Combining dense retrieval with BM25 raised the reported retrieval hit rate from 92% to 97% on 63 questions—a measure of finding evidence, rather than complete answer correctness.

The assistant could expand from search results to sections and full documents, with limits on how much material entered the context. Its median answer time of 38 seconds also shaped the interface: streaming Markdown separately from a completed source card let users see the answer before the entire structured response was ready. The early internal rollout reached 34 users and roughly 900 questions; external release remained planned. These were shared project decisions described by both speakers, rather than individually assigned contributions.

Testing the experience and changing the workflow

Nina’s joint writing develops several concrete positions on how people should work with imperfect models:

• Test the experience before the model. In a prototyping guide co-written with Rémi Delanghe and Peter Vermaercke, she advocates testing interfaces with simulated predictions before building the underlying system. Deliberately varying output quality reveals how users handle mistakes, report problems and recover. Technical feasibility then gets separate acceptance criteria covering data, model performance and operational requirements.

• Let users influence the evidence behind an answer. In 2023 writing with Rémi Delanghe, Nina connects retrieval-augmented generation to a specific interface choice: a legal-translation application let users select historical translation pairs to supply domain-specific context in the prompt. Their recommendations also distinguish situations that benefit from generated answers from those where showing the original source is safer.

• Reconsider the unit of work assigned to an agent. In joint writing with Peter Vermaercke, Nina argues for gradual, monitored adoption while reconsidering the workflow itself. Their customer-support example moves beyond assigning an agent individual tickets: examining many tickets together could reveal a shared outage and change the response to each request. The proposal pairs experimentation with attention to capability limits, accountability and human involvement.