At King, we ingest and analyze around a trillion of game events weekly to enhance gameplay and deliver player experiences to over two hundred million monthly active users. These game events enable us to improve our games via game observability, features delivery, and machine learning.

However, collection and storage of these game events bring a variety of challenges due to its sheer volume and high speed.

To address the challenges, we have developed an in-house product called Kingestor, which allows us to ingest our game events in an effective manner. It processes game events in near real-time with just a ten-minute latency, loading approximately five million events per second. Kingestor ensures data integrity through event reconciliation and deduplication, providing accurate, real-time insights for both business and technical applications. It is a scalable and adaptable product, which is designed for use across the gaming industry, making it easy to implement for businesses handling large-scale data.

Methods to manage data and its authorized usage in Google Cloud Platform, including access control through data governance and enforced security measures:

• Fine grained access control in BigQuery by enforcing IAM policies
• Column level access control by defining taxonomies and policy tags
• Best practices for using taxonomies and policy tags
• Row level access control in BigQuery - Access control restriction by using resource tags.

Best practices like naming conventions, proper data asset descriptions, or ownership tagging, are a must have to ensure proper governance across your data landscape. Yet still they often require manual effort and a trained, knowledgeable user base to put in place, unfortunately leading to those practices not being followed in the majority of cases. The only way to ensure those rules are followed to the letter, is by baking them in as defaults into the user experience of your data platform. HelloFresh is the world's leading meal kit company and global integrated food solutions group, shipping over a billion meals to customers per year. This talk will present HelloFresh's in-house, low-code, config-driven data engineering framework that was constructed to offer data governance and best practices out of the box. You will learn about the architecture around its open source components, and get a demonstration of the user experience designed to enable even less technical data practitioners.

Our presentation will cover the development and adoption of a private AI chat platform within our company. We will discuss the technical architecture, including the integration of multiple large language models (LLMs) and chat engines from various providers. A significant part of the presentation will focus on our user adoption strategy, which includes conducting workshops and identifying internal advocates to promote the technology. We will also highlight how we use this platform as a unified interface for testing new AI experiments, which streamlines development and user feedback. Additionally, we will share our experience in creating separate Python packages for chat engines to ensure scalability and reusability. The aim is to provide a comprehensive view of both the technical and human aspects of implementing AI solutions in a corporate environment.

Discover how an open source project under the Google umbrella leverages GCP key pillar services—advanced LLMs (Gemini), Looker for BI, Cloud Functions, and BigQuery—to drive NLP analytics. Integrating these tools within a React application enables intuitive data querying and visualization. Join this presentation and learn about real-world challenges in production environments.

Every week hundreds of numerical metrics are delivered to decision makers at Allegro. Most of them were calculated within our data ecosystem, some calculated correctly, very few documented or described. We'd like to take you on a quick journey describing what the KPI management process has looked like till now, and how we decided to change it. During our presentation, we'd like to showcase our approach to distributed KPI governance over multiple teams, where metric calculation is orchestrated through Airflow, calculated in GCP and documented using an everything-as-code approach, to end up through CI/CD in one, common data catalog.

This presentation highlights the importance of transaction flagging and outlines the basic business logic. Given the complexity and need for deep business knowledge, we will also introduce an AI Chat Bot integrated with our Confluence as a valuable tool. We will show how we are leveraging vector database search to enhance data retrieval efficiency and accuracy. We will also show how Apache Flink was utilized for real-time data processing and analytics, ensuring timely insights. Together, these technologies are streamlining operations and support informed decision-making.

The session will focus on the concept and importance of data extrapolation, emphasizing its vital role in forecasting and decision-making, especially when historical data is limited for anticipating trends. It will explain the reasons for using extrapolation, discuss potential challenges, and outline effective interpretation methods to enhance prediction accuracy.

Additionally, the session will review common extrapolation techniques, detailing their strengths and limitations. An alternative approach will also be introduced. A practical example will be provided to demonstrate how this alternative method can be applied, offering a clear understanding of its benefits, potential drawbacks, and success measures in real-world forecasting scenarios.

In an evolving regulatory landscape, automation plays a crucial role in ensuring efficiency, consistency, and compliance across the Credit Risk Model Life Cycle.

This session explores how automation enhances key processes from data through model development to monitoring and validation reporting.

We will showcase how automation can streamline the processes around MLC and how AI-driven solutions, including LLMs, can help interpret test results and regulatory changes, providing deeper insights and accelerating decision-making.

By leveraging automation and AI, financial institutions can automate workflows, reduce operational risk, and improve transparency in model governance.

Data lakehouse has recently emerged as a go-to architecture for implementing modern data platforms. It bridges the gap between expansive data lakes and the structured world of data warehouses.

On one hand, its elasticity and modular nature enables enormous flexibility to meet very specific customer needs. On the other, since with great power comes great responsibility, some architectural decisions may significantly impact the overall performance, cost and in the end customer satisfaction.

The current tools available on the market allow us to implement this platform in various models, ranging from fully Open Source to manage SaaS. How can we choose the right tool for a specific case? Is there a universal deployment model for such a platform?

Based on our experience from a few data lakehouse projects we will shed some light on the key aspects to consider while architecting a robust data platform.

Fast Deliveries, a freight forwarding company, faced a challenge when several of their forwarders violated non-competition agreements by working for their competitor, WHILE (names anonymized), and transferring clients during this period. The investigation centered on analyzing data from Trans.eu - a major European logistics platform where freight forwarders post and accept transportation orders, essentially serving as a digital marketplace crucial for day-to-day logistics operations.

Our team was tasked with developing a methodology to identify these fake accounts and connect them to specific former employees. Combining network analysis, geographic patterns, and temporal data allowed us to identify suspicious accounts with high confidence.

Attendees will learn about the specific investigation and gain insights into analytical techniques they can apply to their own data challenges. We will demonstrate our methodology using anonymized data from the actual investigation.

At SiriusXM, we make sure that satellite radio and online radio subscribers receive the optimum recommendations of various music, news, talk & podcast or sport content.

This talk is about the real challenges and best practices of application of the AI, ML and Data Science in the streaming media, focusing on:

• building the right information from raw data
• evaluation of the framework and end-to-end ML platform architecture
• evaluation of the AI and ML algorithms
• quality control, experimentation and measuring the impact of the implemented solutions
• security, compliance and data privacy
• and finally discussing key takeaways from the process of building recommendation systems for media streaming

What makes a great meme? Is it the template? The reference to recent events? Or perhaps sheer luck? Using the image embedding pipeline with the refined Vision Transformer model by Google, we explore the memesphere (yes, it's a word) of Reddit, and it's most popular meme subreddit: r/memes. We brew a recipe for the best memes, by analyzing the upvotes and comments statistics. We determine the most similar memes in terms of content and graphics to establish relations and form clusters segregated by meme templates. Finally, we answer the world-shaking question: What was the best meme of last year?

The future of Artificial Intelligence in software testing: its transformative impact within the telecommunications industry. As AI continues to evolve, organizations are leveraging intelligent solutions to optimize testing processes, enhance product quality, and reduce time-to-market. Drawing from real-world use cases from Ericsson AB, this presentation will dive into how AI is revolutionizing testing methodologies, addressing challenges in AI deployment, and setting the stage for the next leap in testing innovation. Attendees will gain actionable insights into integrating AI into testing pipelines, handling the complexities of large-scale deployments, and overcoming the challenges that come with AI adoption in the software testing space.

We want to showcase that successful implementation of AI solutions (open-weights semantic search for customer care) doesn’t have to be costly and can deliver solid ROI - contrary to the growing sentiment in the media as we seem to be entering the trough of disillusionment (as per Gartner’s hype cycle terminology). In our internal tech talks, our peers were particularly interested in some of the thinking process behind certain decisions that we made. For instance they wanted to know how we evaluated the most suitable model from the MTEB leaderboard, how we organized the embeddings to fit our knowledge base, why we chose this and not a different vector store and more. Interestingly, this implementation and heavy internal promotion among non-tech folks spurred an avalanche of ideas from other departments. We want to spread the experience and believe there’s no better place to do this than at the tech-heavy BigData Technology Warsaw Summit.

As organizations increasingly rely on data to drive decision-making, the ability to quickly access and analyze data has become essential. However, SQL querying remains a technical hurdle for many, slowing down workflows and limiting data access to specialized team members. This session will demonstrate how LLM-based Text-to-SQL tools can close this gap, enabling non-technical team members—such as product managers and business analysts—to generate SQL queries using natural language. In this talk will explore how to implement Text-to-SQL solutions for databases, discuss strategies to improve query accuracy through prompt engineering techniques, and optimize results by incorporating database structure representations. Attendees will walk away with actionable insights on how to empower their teams, streamline query development, and increase overall productivity.

• In today’s fast-changing, competitive business environment, traditional BI dashboards are often too slow to drive real-time decision-making.
• This presentation explores Proactive BI, an approach where AI-driven insights are delivered directly to business users at the right time and in the right context.
• Using a case study from Tpay, we demonstrate how automated BI alerts, integrated with tools like Slack, help teams take immediate action on customer retention, sales anomalies, financial performance, and IT incidents.
• We showcase how AI-powered decision support transforms BI from passive reporting into an active driver of business success.
• Attendees will learn how to bridge the gap between analytics and action, ensuring that business intelligence directly impacts key company goals.

In the fast-paced world of cloud-native applications, ensuring seamless deployment and management of service dependencies is critical. Enter Saxo Blueprint, an innovative solution developed to orchestrate the deployment of essential dependencies such as internal applications, Kafka topics, databases, firewall openings, DNS records, and more. Additionally, a portal based on Spotify Backstage enables developers and other employees to discover and configure services in the company.

This solution enables easier disaster recovery as well as giving us a great way to provide evidence to auditors including approvals.

Restate (https://restate.dev/) is a system built by creators of Apache Flink, and inspired by our work on Apache Flink and Stateful Functions. It is an event-driven system built from the ground up for transactional workflows.
While Flink is built for analytical use cases (think OLAP on streams), Restate can be thought of as the transactional (OLTP) twin counterpart, for transactional use cases like workflows, sagas, AI agents, state machines, and microservice orchestration.

Restate implements an architecture that shares similarities with stream processing, but also takes a very different approach at many layers. Restate sacrifices some of Flink's throughput to achieve very low latency transactions, high availability, serverless deployments, more flexible programming abstraction, more flexible state and queries, and polyglot APIs.

In this talk we contrast durable execution vs. stream processing, their use cases, and how the architecture evolved from Flink over Stateful Functions to Restate.

Recommendation systems are an essential part of most e-commerce industries, often responsible for a significant portion of revenue. However, every branch of this industry has its own set of exceptions and challenges that affect how recommender systems have to be designed. In airlines, these exceptions become extreme as returning visitors become sparse, many purchases are anonymous, and items, such as flight tickets, can be sold at different prices depending on the circumstances. To overcome these challenges, we propose a simple method that utilizes information collected about users and items, omitting the need for extracting user/item embeddings with matrix factorization. Additionally, we will talk about how we used a Feature Store as a foundation for this project and why it could be beneficial to implement it in your Data Science team as well.

Exploration how our system uses a graph-based approach to store transactions and enhance fraud controls with advanced features, boosting the effectiveness of both ML models and static rules. Presentation of key components of the system, including a real-time feature computation service optimized for low latency, a visualization tool for network analysis, and a mechanism for historical feature reconstruction.

The evolution of AI has seen a remarkable transition from standalone language models (LLMs) to compound systems integrating diverse functionalities, culminating in the rise of agentic AI. This talk traces the journey of AI systems, exploring how agentic AI enables autonomous reasoning, planning, and action, making it a pivotal development in solving complex, dynamic problems.
We will dive into the principles of agentic AI, discussing how it works and why it is essential for creating adaptive, task-oriented solutions. The session will then introduce **CrewAI**, an open-source Python package that simplifies the development of intelligent agents. Through a practical use case, participants will learn how to implement their first agent with CrewAI, gaining hands-on insights into leveraging this powerful tool to unlock new possibilities in AI-driven applications.

Over the past 8 years at Uber, I have worked on different aspects of building and scaling Uber's Data Infrastructure (batch and realtime) and gotten to look at microservices and storage systems security in the past 2 years.

This talk aims at enabling security for entire Uber's data infrastructure - a lot of the work which was done over the course of my tenure at Uber. Internally we rely heavily on open source systems for data infrastructure and security (SPIRE and Kerberos are widely adopted).

Motivations for security efforts at Uber include protecting data from cyberattacks, insider threats and staying up to date with compliance requirements.

I'll delve into our journey at Elisa Polystar in deploying large-scale big data platforms for leading telecom operators like Vodafone, TMobile, and Telia. I'll discuss our unique challenges in handling massive volumes of telecom data and achieving real-time analytics. We'll explore the technical architectures we implemented, the tools and technologies we leveraged (such as Hadoop, Spark, and Kafka), and how we customized solutions to meet the specific needs of different markets. The goal is to share practical insights, lessons learned, and best practices that can help others in the industry successfully implement and scale big data solutions.

• Explore how scalable and intelligent data products are designed to transform raw data into actionable insights.
• Uncover the end-to-end architecture enabling seamless data replication, storage, processing, and delivery for business use cases.
• Learn about metadata management, data lineage, and transformation processes that enhance data discoverability and usability.
• Key takeaways from Scania's implementation of MDM and Data Products.

Dive into the practical and strategic considerations when choosing between these two approaches for creating effective AI agents. Prompt engineering has risen as a fast, adaptable, and low-cost way to harness the capabilities of LLMs. However, its performance often correlates directly with the size of the model - larger, more costly models are required to achieve the desired results. This trade-off raises questions about scalability and cost-efficiency, especially for organisations with resource constraints.

On the other hand, fine-tuning offers a path to tailor models for domain-specific tasks or nuanced interactions, delivering consistent performance even with smaller models. While it demands more resources upfront, fine-tuned solutions can lead to significant long-term savings by reducing reliance on oversized models:

• The strengths and limitations of prompt engineering vs. fine-tuning in AI agent development
• Cost implications: why prompt engineering often requires larger, more expensive models to perform well
• Fine-tuning as a solution to achieve domain-specific precision with smaller models
• Case study: the TUI AI travel assistant for the UK market and lessons learned
• A hybrid approach: combining prompt engineering and fine-tuning for best results.

The purpose of presentation is to demonstrate possibilities of Generative AI in the context of business intelligence. I'll focus on Copilot, AI Skills in MS Fabric, and AI/BI Genie in Databricks. The presentation will contain description of each tools, comparison, and the demo at the end. I'll present how to prepare data model and data to make it available for Gen AI. I'll demonstrate how these tools can support data democratization in an organization.

I'm also considering to include Conversation Analytics from GCP

This presentation delves into creating a real-time analytics platform by leveraging cost-effective Change Data Capture (CDC) tools like Debezium for seamless data ingestion from sources such as Oracle into Kafka. We’ll explore how to build a resilient data lake and data mesh architecture using Apache Flink, ensuring data loss prevention, point-in-time recovery, and robust schema evolution to support agile data integration. Participants will learn best practices for establishing a scalable, real-time data pipeline that balances performance, reliability, and flexibility, enabling efficient analytics and decision-making.

Full info

I want to show you the principles behind the most popular DBMS, enumerating non-trivial use cases like self-containing dynamic reports, WebAssembly support and pure serverless DB hosting. While SQLite works well with aggregated datasets, DuckDB, its younger cousin, focuses on full OLAP support, allowing processing gigabytes of data in no time on low-end boxes (or even laptops). We will browse various useful features and interfaces in DuckDB, emphasizing scenarios that anyone can implement in their daily work.

Full info

In this workshop, we’ll introduce the key components of a multitier architecture designed to scale and streamline LLM productization at Team Internet—a global leader in online presence and advertising, serving millions of customers worldwide. For us, scalability and speed are critical to delivering high-performance services, including LLM applications.

Through hands-on coding exercises and real-world use cases from the domain name industry, we'll demonstrate how standardization enhances flexibility and accelerates development.
By the end of the session, you’ll know the keybuilding blocks that will help you efficiently building and scaling LLM applications in production.

Full info