Spark Sessions 007
The Spark Sessions are all about about igniting your creativity and passion through inspiring talks and discussions!
Spark Sessions 007
The Spark Sessions are a recurring event of the Data Science Initiative at the Faculty of Computer and Information Science. They are designed to facilitate the exchange of ideas and inspiration. The event consists of several snappy (5-7 min) presentations, followed by a casual gathering with refreshments:
1. Building a Movie Recommendation Agent by Rok Novosel (Member of Technical Staff, Poolside)
I'll talk about how I built a movie recommendation LLM agent (movieagent.io) to solve Friday movie nights for me and my wife.
I'll go through the following:
- How to design multi-agent architectures for recommendations
- How to improve retrieval with synthetic data
- How to design the user experience around agents
- How to evaluate an agent-driven recommendation system
2. Lestvica embeddingov za slovenščino (LES) by Neli Čatar (Student, UL FRI and Data Scientist, Valira AI)
We will be presenting Lestvica embeddingov za slovenščino (LES), which is essentially a Slovene version of the MTEB benchmark for embedding models with tasks including classification, clustering and retrieval. In the talk, we will discuss the data collection and the benchmark implementation together with some insights into the best performing models.
3. Decisions, Not Predictions: ML Won't Cut Your Energy Bill by Nevena Pivač (Researcher, Abelium)
We often say data science is statistics, algebra, and optimization -- but in practice, optimization usually means minimizing a loss function and not much more. There's a whole world beyond that. Energy systems are a great example: when should the battery charge? How do you manage large systems? The answer doesn't require a forecast, it requires a decision. In this talk, we explain how to handle energy scheduling, and explore when and why a MILP model beats prediction-based methods.
4. From Billions of Points to Meaningful Maps by Matic Murko Drozdek (Machine Learning Engineer, Flai)
Every year, LiDAR sensors scan hundreds of thousands of square kilometers, producing enormous 3D point clouds. But raw point clouds are only the starting point: their real value comes from understanding what is in them. This talk shows how Flai uses 3D computer vision and deep learning to classify buildings, roads, vegetation, and power infrastructure in large LiDAR datasets, turning raw scans into actionable geospatial insight.
5. Amazonia Archaeology Discovery by Maja Somrak (Data Scientist, ML Engineer)
This is a self-initiated project that achieved 3rd place in the OpenAI to Z Challenge (2025). The project integrates satellite imagery analysis and domain knowledge to support the discovery and interpretation of archaeological structures in the Amazon rainforest.
6. D.A.T.E.: Scalable document anonymization platform by Gal Petkovšek (Data Scientist, Medius)
In the era of stringent data privacy, moving from simple entity detection to a production-ready anonymization system remains a significant hurdle for enterprises. This talk introduces D.A.T.E. (Data Anonymization Tool for Enterprise), a robust platform designed to automate the anonymization and pseudonymization of sensitive information within complex, unstructured datasets and image formats.
Built upon the Presidio framework, D.A.T.E. employs a sophisticated hybrid detection engine. By ...
7. Snacks and drinks and discussions
The Spark Sessions are all about about igniting your creativity and passion through inspiring talks and discussions!
Spark Sessions 007
The Spark Sessions are a recurring event of the Data Science Initiative at the Faculty of Computer and Information Science. They are designed to facilitate the exchange of ideas and inspiration. The event consists of several snappy (5-7 min) presentations, followed by a casual gathering with refreshments:
1. Building a Movie Recommendation Agent by Rok Novosel (Member of Technical Staff, Poolside)
I'll talk about how I built a movie recommendation LLM agent (movieagent.io) to solve Friday movie nights for me and my wife.
I'll go through the following:
- How to design multi-agent architectures for recommendations
- How to improve retrieval with synthetic data
- How to design the user experience around agents
- How to evaluate an agent-driven recommendation system
2. Lestvica embeddingov za slovenščino (LES) by Neli Čatar (Student, UL FRI and Data Scientist, Valira AI)
We will be presenting Lestvica embeddingov za slovenščino (LES), which is essentially a Slovene version of the MTEB benchmark for embedding models with tasks including classification, clustering and retrieval. In the talk, we will discuss the data collection and the benchmark implementation together with some insights into the best performing models.
3. Decisions, Not Predictions: ML Won't Cut Your Energy Bill by Nevena Pivač (Researcher, Abelium)
We often say data science is statistics, algebra, and optimization -- but in practice, optimization usually means minimizing a loss function and not much more. There's a whole world beyond that. Energy systems are a great example: when should the battery charge? How do you manage large systems? The answer doesn't require a forecast, it requires a decision. In this talk, we explain how to handle energy scheduling, and explore when and why a MILP model beats prediction-based methods.
4. From Billions of Points to Meaningful Maps by Matic Murko Drozdek (Machine Learning Engineer, Flai)
Every year, LiDAR sensors scan hundreds of thousands of square kilometers, producing enormous 3D point clouds. But raw point clouds are only the starting point: their real value comes from understanding what is in them. This talk shows how Flai uses 3D computer vision and deep learning to classify buildings, roads, vegetation, and power infrastructure in large LiDAR datasets, turning raw scans into actionable geospatial insight.
5. Amazonia Archaeology Discovery by Maja Somrak (Data Scientist, ML Engineer)
This is a self-initiated project that achieved 3rd place in the OpenAI to Z Challenge (2025). The project integrates satellite imagery analysis and domain knowledge to support the discovery and interpretation of archaeological structures in the Amazon rainforest.
6. D.A.T.E.: Scalable document anonymization platform by Gal Petkovšek (Data Scientist, Medius)
In the era of stringent data privacy, moving from simple entity detection to a production-ready anonymization system remains a significant hurdle for enterprises. This talk introduces D.A.T.E. (Data Anonymization Tool for Enterprise), a robust platform designed to automate the anonymization and pseudonymization of sensitive information within complex, unstructured datasets and image formats.
Built upon the Presidio framework, D.A.T.E. employs a sophisticated hybrid detection engine. By ...
7. Snacks and drinks and discussions
Partnering with Forum podatkovne analitike.
Nearby parking options
Good to know
Highlights
- 2 hours
- In person
Location
Faculty of Computer and Information Science
113 Večna pot
1000 Ljubljana
How do you want to get there?
