Big Techday 26 | Talks by TNG

Speaker: Jonathan Rohland

Jonathan Rohland, Principal Consultant at TNG, demonstrates how agentic coding tools such as Cursor can be used to automate recurring tasks in the daily work of a product owner. In the talk, he details the challenges he faces in his product owner role in a 40-person project spanning 200 stories and how to accelerate their resolution with the help of AI. Using practical examples ranging from Jira reporting and meeting follow-ups to knowledge management, he shows how deterministic data pipelines can be combined with targeted AI-powered processing. He highlights how information can be distilled more effectively, how dependencies can be updated efficiently, and how decisions are easier to trace. At the same time, it is underlined that human judgement remains essential and that AI is most effective when augmenting existing workflows rather than replacing them.

Speakers: Dr. Florian Gather, Tobias Weyer & Fabian Lübbe

Dr. Florian Gather, Partner at TNG, and Consultants Tobias Weyer and Fabian Lübbe detail how they built a cost-effective humanoid robot based on Berkeley's open-source DIY design. To avoid the high costs of commercial systems, the team used 3D printing for more than 120 parts. The speakers discuss how they managed the inherent hardware challenges of this approach, such as cable strain relief, encoder/sensor calibration, and repeated mechanical failures. They also explain how they used reinforcement learning to train their own walking policy in simulation and how ROS can be leveraged to deploy learned behaviors on physical hardware. Overall, this session provides a practical perspective on both the opportunities and the engineering effort involved in open-source humanoid robotics.

Speakers: Lukas Hauser, Dr. Moritz Maus & Dr. Miguel Obando Leitón

In this talk, Dr. Miguel Obando Leitón, Dr. Moritz Maus, and Lukas Hauser present an AI-assisted approach for modernizing business-critical COBOL systems. They share insights from a complex project that involved migrating code from COBOL to idiomatic Java. The speakers explain why directly translating legacy code with large language models is rarely sufficient, and highlight challenges such as implicit logic, decades of accumulated complexity, and unreliable documentation. They also introduce a six-step modernization process that combines deterministic methods with carefully targeted use of LLMs and validates results through automated behavior-based testing. Together with incremental migration strategies and modern cloud architectures, this approach enables the low-risk transformation of legacy applications that are still in use today.

Speakers: Mathias Burger & Áron Erdelyi

In this talk, Mathias Burger and Áron Erdelyi explain the core concepts of the Model Context Protocol (MCP) and trace its history from a pragmatic interface to becoming the de facto standard for tool calling and reusable workflows in AI systems. They discuss the challenges that have come with MCP's rapid growth, for example fragmented implementations and increasing context bloat, and emphasize important security considerations, such as prompt injection, tool poisoning, and the handling of sensitive data. The speakers also focus on adjacent ideas and protocols: MCP Apps and A2UI for generative user interfaces, as well as A2A for long-running, multi-agent interactions. After the talk, you will have an overview of different agentic protocols, how they fit into the larger AI landscape, and how these efforts are shaping the next generation of agentic systems.

Speakers: Daniel Klingmann & Max Dietsch

In this talk, Senior Consultant Daniel Klingmann and Software Consultant Max Dietsch compare artificial intelligence with the human brain. To make complex ideas easy to grasp for people not familiar with AI, they use intuitive, everyday analogies that simplify key concepts. Their goal is to build an intuitive understanding of AI, highlight similarities and contrast the differences with human cognition.

Drawing on insights from neuroscience and AI research, the speakers illustrate how foundational architectural concepts, such as functional specialization and sparse activation, are directly mirrored in modern systems like Mixture-of-Experts (MoE) models. They go on to highlight striking parallels in how biological and artificial minds learn. Finally, they dive into how thoughts actually form – a process we still barely understand – and what's really going on beneath the surface during inference and internal representation.

Speakers: Thomas Endres & Jonas Mayer

In this talk, Thomas Endres and Jonas Mayer demonstrate how game engines can be used to efficiently train and evaluate robot foundation models for new control tasks. Using Counter-Strike as a reproducible test environment, they show how core robotics challenges - including continuous action control, real-time inference, and imitation learning - can be investigated under realistic conditions.

The speakers explain how they adapted NVIDIA's GR00T-1.5 model for this task, addressing challenges such as inference latency, discrete input signals, and discontinuous movement. They discuss the role of high-quality demonstration data, analog control inputs, and techniques such as real-time action chunking in achieving stable control. The talk illustrates how these engineering principles transfer to industrial robotics applications that rely on simulation-based training and learning-based real-time control.