AI Day Schedule

AI Agents are the next big thing everyone has been talking about. They are expected to revolutionize various industries by automating routine tasks, mission critical business workflows, enhancing productivity, and enabling humans to focus on creative and strategic work. Of course, you can apply them to your everyday coding tasks as well.

In this hands-on workshop, we will learn how to develop a custom observability solution tailored for AI agents. From logs and metrics to guardrail checks and cost tracking, we’ll discover how to monitor crucial elements like accuracy, latency, and environmental impact. Walk away with your own tool and best practices to ensure AI agents are both well-observed and well-behaved.

We built a cutting-edge speech-to-text model that outperformed solutions from industry leaders like Microsoft, Google, and OpenAI. This is the story of how we identified a market gap, defined what makes "the best model," and turned our vision into a successful business.

Fairness and safety are fundamental criteria for building trustworthy and high-quality AI systems, whether they are credit scoring models, hiring assistants, or healthcare chatbots. But what does it truly mean for an AI system to be fair and safe? In this talk, I will explore the potential risks and challenges associated with these principles and introduce various approaches and techniques for evaluating AI systems. The discussion will center on applications powered by Large Language Models (LLMs).

Deploying PyTorch models into production requires balancing performance, hardware efficiency, and cost optimization. Developers face challenges such as GPU memory constraints during training, suboptimal inference latency, and escalating cloud infrastructure expenses. For instance, modern models like LLMs demand specialized strategies to reduce GPU memory footprint while maintaining throughput.

Some of the latest big evolutionary steps in generative AI has been models that support function calling and “agentic” capabilities. This is provides generative models with “tools” that allow them to go beyond generating outputs for simple queries, and start planning the best way to solve complex queries. In this talk, we’ll be diving into using vector databases as the backbone for these types of complex AI architectures, both serving as knowledge bases, and memory.

This talk presents a retrieval-augmented generation system tailored for emotion analysis in Nigerian Afrobeats lyrics. Integrating supplementary dictionaries with a hierarchical classifier can enable the system retrieve culturally relevant context to refine emotion predictions. We discuss the architecture, implementation challenges and potential for addressing code-mixing and linguistic nuances in low-resource settings.

AI is evolving from passive tools to autonomous agents, driving the rise of agentic workflows that can plan, execute, and optimize tasks with minimal human input. This session will explore how large language models and multi-agent collaboration power these systems, enhancing efficiency and innovation. Attendees will gain insights into real-world applications, challenges, and the future of autonomous AI systems, uncovering how agentic workflows are transforming industries.

In an era dominated by data, businesses struggle with processing diverse, unstructured information across systems. This research presents an AI-powered pipeline addressing product matching challenges in retail and e-commerce. Our solution combines traditional matching algorithms with deep learning through a five-step process. This approach minimizes manual intervention while improving accuracy and efficiency.

The more users your platform attracts, the more unwanted attention you'll get from people looking to game your system. While every product is unique, the journey of tackling these bad actors tends to follow similar patterns across companies. In this talk, I'll walk you through the three stages of platform protection that I've witnessed firsthand, and how to level up your safety game using the data you have.

Scoris.lt utilized Large Language Models (LLMs) to address the challenge of improving SEO performance by generating financial descriptions for Lithuanian companies. The case study highlights the innovative application of LLMs and custom translation models to create high-quality, multilingual content at scale.

Modern work demands constant context-switching—emails, notes, meetings, and tasks pile up, leaving us overwhelmed. This talk introduces a <b>slow productivity AI</b> approach, inspired by <b>Cal Newport</b>, that leverages <b>offline, open-source automation</b> using <b>Hugging Face, n8n, and Obsidian</b>. By structuring knowledge into meaningful tasks <b>without disrupting deep work</b>, we can create a <b>sustainable, low-distraction workflow</b>—working smarter, not just faster.

AI agents are becoming fundamental building blocks of modern applications, thanks to their capability of understanding context with the help of the LLMs, but with added capability to interact with systems, files. AI agents that can meaningfully interact across platforms while maintaining context aren't just for large tech companies anymore as we have a number of open source AI Agent frameworks now which developers can leverage. This beginner-friendly, hands-on tutorial shows the process of creating autonomo

This session provides a concise overview of EU AI Act compliance tailored for Python developers building AI systems. We'll explore key compliance areas, including risk management, robust model testing, cybersecurity, and explainability, highlighting practical tools and popular Python libraries.

AI agents are transforming the way we create applications. However, developing multi-agent applications can often feel complex and time-consuming. LangFlow simplifies this process by offering an intuitive, easy-to-use interface for building AI-driven solutions.

This talk charts the evolution of Artificial Intelligence through the dual lenses of data and models, tracing AI’s journey from early symbolic systems to today’s advanced data-driven techniques. Attendees will learn how the interplay of ever-growing datasets and increasingly sophisticated model architectures has powered major breakthroughs, transforming AI from theoretical curiosity to a global catalyst for innovation.

Take your web apps to the next level by combining FastAPI's speed and efficiency with LamaIndex's AI-driven search engine. This talk will show you how to create powerful search-as-a-service APIs. Unlock new possibilities for your users with advanced search capabilities.

How can machine learning enhance biomedical image analysis? This talk explores the potential of Python and PyTorch in automating artifact and damage segmentation. From data preprocessing to clustering-based label classification and deep learning-driven segmentation, key techniques will be discussed, including the use of Convolutional Neural Network architectures. The session will also cover performance evaluation and insights into advancing biomedical imaging with AI-driven solutions.

As artificial intelligence continues to evolve, there is a growing need for project maintainers to leverage on AI to foster diverse and inclusive collaborations. In this session, I will discuss importance of artificial intelligence in open-source project management, actionable strategies to maximize diverse talents and promote transparency, tools and technologies available for project maintainers, and practical ways to leverage these technologies to make a meaningful impact within open-source communities.

AI-driven code generation can transform software development in regulated sectors like banking and insurance - but only if implemented securely and responsibly. In this talk, we’ll explore how to harness tools like GitHub Copilot and ChatGPT to boost productivity while ensuring compliance. Attendees will learn key considerations, best practices, and practical insights to keep code generation both efficient and fully auditable.

Unlock sensitive data potential with anonymization! Learn how Python, diffusion models, and Named Entity Recognition (NER) empower institutions to anonymize PII in financial documents, replacing it with synthetic stand-ins. Discover open-source, self-hosted tools to ensure privacy while unleashing data's full power.

Discover how EGTL (Extract, Generate, Transfer, Load) extends traditional ETL by adding a “generate” step powered by GenAI. In this talk, I’ll demonstrate how Python pipelines on top of data warehouse can automatically extract data, generate new insights, and deliver optimized transformations. We’ll explore practical workflows, real-world use cases, and best practices—equipping you to apply EGTL in your own data projects.

What is the way from prove of concept to big production solutions for GenAI application? How to make it scalable and make one release by 7 sprints?

You have probably heard of vector databases while learning about embeddings used in deep learning models or while exploring Retrieval Augmented Generation(RAG). In that case, you might be wondering why not just use SQL databases or NoSQL databases. In this talk you will learn what embeddings are, and what vector databases are, why they are useful and when to use them. Expect to understand why vector databases are preferred for certain applications over SQL and NoSQL databases.

What can go wrong with tokenizer encodings? Everything! I will share my experience of understanding, misunderstanding, and ultimately learning to work with tokenization in LLMs. I will discuss what surprisal is, its relevance to my research, and its connection to tokenization. The talk will include various examples illustrating how misunderstandings of tokenization can arise, as well as strategies for debugging and preventing these issues.

The Model Context Protocol (MCP) is an emerging standard that enables structured data provisioning for LLMs and AI agents. However, the current data discovery mechanism in MCP is static. This limits the AI’s ability to dynamically assess the utility, relevance, and efficiency of data tool calls in real time. Here I present an enhancement to MCP "tool discovery" that introduces dynamic data descriptions, allowing LLM to be better informed.

.