Arrive and connect with others.

Welcome to the dida conference.

In this talk, we present miniCOIL - our attempt to make a sparse neural retrieval model as it should be - combining the benefits of dense and lexical retrieval without propagating their drawbacks. We will share how to design and train a lightweight model that is performant on out-of-domain data and demonstrate its capabilities.

Contemporary Large Language Models (LLMs) constitute a powerful but often opaque enabler technology. This talk explores how tools from Explainable AI (XAI) and mechanistic interpretability can help us uncover the roles of learned structures - such as attention heads or knowledge neurons - enabling more reliable and actionable interactions with AI. We demonstrate how local explainability reveals when an attention head actively contributes to inference and where it draws information from, while global explainability and mechanistic tools uncover its broader function. Using these insights, we introduce computationally efficient methods to detect when an LLM "lies" relative to its own knowledge, identify conflicts between context and parametric memory, and annotate each generated token with measurable reliability scores. Our techniques operate with minimal computational overhead, making real-time, citation-worthy, and controllable LLM outputs feasible in cost-sensitive applications. These advancements push LLMs toward greater transparency, reliability, and practical integration in real-world systems.

AI-powered data extraction works - until it doesn’t. When handling structured tables in invoices, orders, or financial documents, we expect OCR, LLMs, and Vision AI to extract data reliably. However, complex documents - e.g. nested tables, irregular structures, and edge cases - pose real challenges for document data extraction AI models. With our solution smartextract, we tackled a real-world customer challenge: automating order entry from complex order documents and tables for a German shoe retailer: OCR and text-based LLMs struggled, Vision LLMs were inconsistent. Only extensive customization could solve the appearing problems - including segmentation, few-shot prompting, fine-tuning, and even the possibility of training a custom computer vision model. In this talk, we will show why standard AI models struggle with complex tables and demonstrate in which cases segmentation helps. Further, we will show benchmarks of commercial vs. open-source models and discuss the trade-offs between OCR, LLMs, and computer vision models.

In the age of digital content, music and sound are crucial elements that enhance storytelling, engagement, and emotional impact. Yet, finding the perfect soundtrack for a video - one that aligns seamlessly with its mood and pacing - remains a challenge for many creators. Mirelo AI is changing this with cutting-edge generative foundation models for AI-driven music and sound design. In this talk, we will introduce our approach to video-aware music generation. Our technology enables users to upload a video, and our AI dynamically composes original music and synchronized sound effects tailored specifically to its content. This is a game-changer for content creators, advertisement agencies, game developers, and everyday users looking to elevate their projects with custom-generated audio.

This talk covers strategies applied in bringing a real-time license plate matching service into production. Based on the project requirements, we discuss the chosen ML model architecture and training procedure, implemented software architecture components, as well as tools and infrastructure for the deployment. Particular focus will be on the deployed microservices architecture and model serving using NVIDIA Triton.

With confidential computing, data can remain protected or encrypted even during processing. Since the introduction of the Nvidia H100, this technology can also be applied to AI. When used correctly, it enables the creation of AI-as-a-Service (such as ChatGPT) where users do not need to trust the provider regarding data security and privacy. This makes it possible to process sensitive with GenAI services. The presentation will explain the fundamentals of the technology and provide an overview of use cases and existing deployments.

In this talk, we present a general-purpose model architecture that is suitable for detecting and counting subtle events in videos. It combines a convolutional neural network (CNN), a recurrent neural network (RNN) and a YOLO-type head in the time domain. The model can be trained end-to-end with time-resolved labels only, without the need for customization or domain-specific knowledge. We discuss possible use cases and show experimental results both on synthetic datasets and on real world datasets.

The posters will be announced later. If you would like to present a poster, please send us a message to conference@dida.do.