Bernhard Merkle
Gen AI, AI, ML, Cloud, XOps, Technology Scout, MS MVP
Freiburg, Germany
Actions
Bernhard Merkle is the Head of AI Solution Strategy at blockbrain.ai, where he advises C-level executives and shapes the company's AI product strategy. Passionate about real-world impact, Bernhard focuses on delivering production-grade agentic AI systems for enterprise customers rather than mere proofs of concept. He previously served as blockbrain.ai's Head of Forward Deployed Engineering, successfully implementing complex AI use cases.
A former Technology Scout at SICK AG and Microsoft MVP for AI, Bernhard combines deep engineering expertise with commercial vision. He is a passionate Python and TypeScript developer (started with C++), an author of technical articles, and a university lecturer. As a recognized thought leader in the space, he frequently speaks at international conferences such as AgentCon, AI Summit, QCon, PyConUK, ACCU, OOP, OOPSLA and various meetups e.g. GitHubNext and Microsoft events.
Links
Area of Expertise
Topics
TypeAgent: From RAG to Structured RAG: Building Long-Context, Precise, and Fast AI Systems
Retrieval-Augmented Generation (RAG) has become the de-facto method for grounding Large Language Models in enterprise data. But classic RAG quickly reaches its limits: conversations drift, pronouns lose their meaning, long-term state becomes ambiguous, and complex multi-step queries often fail. In this talk, we introduce Structured RAG, a next-generation approach implemented in TypeAgent, an open-source framework originally prototyped by Guido van Rossum. Structured RAG keeps conversation context well-organized, reduces hallucinations, improves retrieval precision, and allows AI agents to handle multi-turn dialogues and complex instructions—while remaining fast and highly scalable.
We walk through the core architecture of Structured RAG, including typed conversations, semantic memory layers, indexing strategies, and automatic disambiguation of historical references. Live demos show how TypeAgent enables long-running AI agents that remain accurate over dozens of turns. Attendees will gain a deep understanding of how to move beyond vanilla RAG and how to incorporate Typed Conversations into real-world Python systems.
I am an active contributor to TypeAgent, an open-source framework initiated by Guido van Rossum for building typed conversational agents and Structured RAG systems. I works on large-scale AI deployments, document understanding pipelines, and knowledge-graph-driven retrieval systems in industrial environments.
The Rise of Agentic Workflows: AI That Understands, Decides & Codes
Software development is entering a new era where natural language becomes a first-class interface for automation. In this session, we explore Agentic Workflows, a cutting-edge research project from GitHub Next that enables developers to describe complex software tasks in plain language — and let intelligent agents execute them. These workflows understand context, make decisions, orchestrate tools, and perform meaningful actions across repositories, going far beyond traditional CI/CD pipelines.
We provide a deep technical dive into how Agentic Workflows operate: the execution engine, decision-making loops, context grounding, tool calling, safety constraints, and how agents interact with GitHub APIs. Through live examples, we demonstrate how natural-language instructions can trigger autonomous refactoring, dependency updates, code generation, documentation improvements, and multi-step engineering tasks.
Participants will gain insight into the emerging discipline of Agentic Engineering: designing systems where agents collaborate with developers, support daily workflows, and automate engineering tasks that previously required manual effort. We discuss practical use cases, limitations, and the future potential for A2A (Agent-to-Agent) architectures running across development teams and repositories.
This talk gives a first-hand look into the next frontier of software engineering — where code, automation, and AI converge into intelligent, natural-language-driven workflows. I am a contributor to the https://github.com/githubnext/gh-aw project, however in agentic time you mainly submit issues (see bug tracker) and the agents fix/implement new features. Looking into the future, we will have to change the contributor metrics (currently based on LOC)
Moving Beyond the Sandbox: Scaling AI Agents from POC to Production
Building an AI agent proof-of-concept (POC) is easier than ever, but transitioning that prototype into a reliable, enterprise-ready production system introduces a completely new set of challenges. This talk breaks down the critical bridge between experimentation and production. We will explore how to transition from brittle prompt engineering to robust agentic architectures, implement rigorous evaluation frameworks to measure accuracy, and establish Azure-backed guardrails for security and cost control. Attendees will leave with a practical roadmap and real-world design patterns to transform their raw POCs into resilient, scalable AI agents that deliver measurable business value.
Industrial Sensor Agents: From RAG to Agentic A2A
Industrial sensors are becoming more capable, more configurable, and more software-defined. Yet developer support for programming these devices often lags behind their technical complexity. In this talk, we share how SICK is building an agent-powered developer experience for a programmable industrial sensor platform using a domain-specific LUA-based framework.
We walk through our journey from an initial RAG-based GitHub Copilot extension—designed to provide code completions and contextual guidance for our APIs—to a conversational developer assistant supported by a custom Bot and an MCP (Model Context Protocol) server. Building on these foundations, we are now aiming to transition to a fully agentic approach, moving toward an A2A (Agent-to-Agent) architecture capable of richer reasoning, tool use, and autonomous workflows.
Participants will learn how structured RAG, domain-specific knowledge models, and MCP-based tooling can evolve into multi-agent systems that provide deep, context-aware support for engineering teams. We highlight our practical lessons, pitfalls, and architectural decisions—showing how agentic technologies can accelerate onboarding, reduce development errors, and unlock a new generation of intelligent support for industrial sensor developers.
From Python to Power: Unlocking AI Acceleration with Mojo
Discover how Mojo bridges the gap between Python’s simplicity and performance/power needed for AI. This talk explores Mojo’s powerful features, ahead-of-time compilation, and seamless integration with Python—empowering developers to write high-performance code without leaving the Python ecosystem behind. Python just got supercharged.
Python is beloved for its simplicity and rich AI ecosystem—but when it comes to performance, especially for GPU and AI workloads, it often hits a wall. Mojo is here to change that. Billed as a superset of Python with systems-level speed, Mojo lets you write high-performance, low-level code with Pythonic elegance.
In this talk, we’ll explore how Mojo blends the ease of Python with the power of C, unlocking native speed, hardware-level control, and GPU acceleration—all within a familiar syntax. Mojo compiles to machine code using MLIR, and interoperates seamlessly with Python: it’s a game-changer for AI, ML, and systems programming.
Through hands-on examples, we’ll show how Mojo enables to write blazing-fast functions—all in the same environment where you'd write your data science experiments. Whether you're building custom operators for AI frameworks or optimizing critical compute paths, Mojo lets you stay productive and performant.
Agents at scale
As AI agents transition from simple chatbots to autonomous workplace assistants, organizations face a critical challenge: how to scale development without losing control. A fragmented toolset creates security gaps and operational chaos.
This session explores the necessity of a unified agent ecosystem. Using the Microsoft Agent Platform as a blueprint, we will examine how to effectively build intelligent agents, centrally govern their orchestration, and secure them with enterprise-grade guardrails. Learn how to balance autonomous agent capabilities with strict compliance and security standards.
Bernhard Merkle
Gen AI, AI, ML, Cloud, XOps, Technology Scout, MS MVP
Freiburg, Germany
Links
Actions
Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.
Jump to top