Demetrios chats with Arpita Vats about how LLMs are shaking up recommender systems. Instead of relying on hand-crafted features and rigid user clusters, LLMs can read between the lines—spotting patterns in user behavior and content like a human would. They cover the perks (less manual setup, smarter insights) and the pain points (latency, high costs), plus how mixing models might be the sweet spot. From timing content perfectly to knowing when traditional methods still win, this episode pulls back the curtain on the future of recommendations.

A new paradigm called the "Agentic Cloud," driven by Generative AI, is set to disrupt the dominance of cloud hyperscalers. While today's giants like AWS and Azure rely on vast but complex service catalogs, this emerging model uses intelligent agents to translate high-level user intentions into fully provisioned and autonomously managed infrastructure. This approach threatens to commoditize the underlying cloud platforms, shifting the primary value from the provider's software ecosystem to the intelligence of the agent itself. Despite significant technical challenges in reliability and security, the Agentic Cloud promises to democratize access to elite infrastructure, enabling a new wave of innovation by shifting the industry from a catalog-driven to an intelligence-driven model.

Agents in Production [Podcast Limited Series] Episode Nine – Training LLMs, Picking the Right Models, and GPU Headaches Paul van der Boor and Zulkuf Genc from Prosus join Demetrios to talk about what it really takes to get AI agents running in production. From building solid eval sets to juggling GPU logistics and figuring out which models are worth using (and when), they share hard-won lessons from the front lines. If you're working with LLMs at scale—or thinking about it—this one’s for you.

What happens when you empower AI agents to design, configure, and deploy other agents? At Hypermode, we put this question to the test by developing Concierge—an agent that acts as both architect and orchestrator, assembling custom agent workflows on demand. In this session, I’ll share the technical journey behind building Concierge, our “agent that builds agents,” and how it’s reshaping the way teams approach automation and task completion. Key topics will include: The architecture and design patterns enabling agent creation How Concierge leverages natural language and user intent to assemble tailored agent teams Real-world challenges: managing reliability, evaluation, and guardrails when agents are in charge Lessons learned from deploying agent-built agents in production environments The future of agentic systems: towards self-improving, self-deploying AI teams

In this talk we will bring to light the open secret in the AI community: most agents don't work reliably. We'll explore the most common ways agents fail, highlighting how fundamental issues with the model often can't be overcome with prompting. If this is true, then why aren't we correcting the paths in the model? Reinforcement learning offers the most promising path to reliable agents. Designing reward signals is the future of agentic development. In the next few years we will transition from agents that are programed deterministically, to agents that are taught interactively. We don't need to be stuck in the ice age of frozen models, we can take our agents to the next level of success.

Synthetic data plays an important role in the news ecosystem. Publishers are now monetizing a synthetic version of their data to help feed news-hungry agents in the wild. We discuss how grounded synthetic news data not only protects publishers against copyright infringement, but also reduces hallucination rates for broad agents built to use hundreds of tools. As agents become better and better generalists, the data that they retrieve via tool-use needs to be packed up and ""Context Engineered"" for quality and ease of consumption. The ancient adage was never more relevant ""Quality in -> Quality out"". Enter the world's largest news knowledge graph. A perfectly searchable, highly accurate, news context delivery machine - geared for high-stakes decision making agentic tasks far and wide. Some tasks include fact-checking, geopolitical risk analysis, event forecasts for prediction markets, and much much more.

Agents are only as useful as the data they can access. EnrichMCP turns your existing data models, like SQLAlchemy schemas, into an agent-ready MCP server. It exposes type-checked, callable methods that agents can discover, reason about, and invoke directly. In this session, we’ll connect EnrichMCP to a live database, run real agent queries, and walk through how it builds a semantic interface over your data. We’ll cover relationship navigation (like user to orders to products), how input and output are validated with Pydantic, and how to extend the server with custom logic or non-SQL sources. Finally, we’ll discuss performance, security, and how to bring this pattern into production.

Evaluating LLM performance is vital to successfully deploying AI to production settings. Unlike regular machine learning where you can measure accuracy or error rates, with text generation you're dealing with something much more subjective, and need to find ways to quantify the quality. As we combine LLMs together and add other tools in the agentic context, this becomes even more challenging, requiring robust evaluation techniques. In this talk I propose an approach to this evaluation that borrows from academic evaluation - namely, creating clear rubrics that spell out what success looks like in as close to an objective fashion as possible. Armed with these, we can deploy additional tested LLMs to conduct evaluation. The result is highly efficient and solves much of the evaluation dilemma, although there are still gaps that I will also discuss. (This is an adaptation of an article I wrote: https://towardsdatascience.com/evaluating-llms-for-inference-or-lessons-from-teaching-for-machine-learning)