Show Notes
Summary In this episode Ofer Mendelevitch shares what it really takes to evaluate RAG systems in production when your source data is incomplete, constantly changing, or difficult to validate against a clean ground truth. He explores how RAG has evolved from simple “chat with your PDF” demos into enterprise-grade retrieval systems that require robust ingestion pipelines, hybrid search, reranking, multimodal support, access controls, and refresh strategies for large and dynamic document collections. Ofer also explains why evaluation becomes one of the hardest parts of the stack, particularly when retrieval quality, generation quality, and system behavior can all shift as you change chunking, models, prompts, or ranking components. He digs into practical approaches for offline and online evaluation, including synthetic question generation, sampling production traffic, and LLM-as-a-judge techniques such as UMBRELA and Auto Nuggetizer. Along the way, he discusses the rise of agentic RAG, context engineering, MCP-connected tools, text-to-SQL systems, and the growing role of coding agents in building and maintaining AI pipelines. Ofer closes with his perspective on where retrieval, memory, multimodal systems, and agent orchestration are headed next, along with a reminder that the biggest gaps in AI today are often not just technical, but educational and organizational as well. Announcements Hello and welcome to the AI Engineering Podcast, your guide to the fast-moving world of building scalable and maintainable AI systems Your host is Tobias Macey and today I'm interviewing Ofer Mendelevitch about how to evaluate your RAG systems when you have incomplete or changing data to grade against Interview Introduction How did you get involved in machine learning? RAG was one of the first concepts to gain traction in the early explosion of LLMs and their application to real-world problems. How would you characterize its role in the current landscape? RAG is conceptually straight-forward, but can be quite complex to do well. What are some of the common pitfalls that teams encounter when bringing a retrieval-oriented LLM service into production? What are some of the predominant methods to avoid regressions or increased error rates as the data, code, and models change in a given deployment?There are numerous software projects and companies that offer various forms of evaluation/validation for RAG workloads. What are the factors that teams should be evaluating when selecting which one(s) to adopt?
Especially when working with data that is subject to change (user-contributed content, changing products, news-based data, etc.), what are the techniques that teams can use to be sure that the generated responses based on that information is accurate? As models have become more powerful the idea of "agentic RAG" started circulating. How is that different from the first iteration of RAG systems?What new pressures and challenges does that prevent? What are the ways that it can improve or simplify a content-based generative system?
How do you think about the differentiation of RAG and "context engineering"? What are the most interesting, innovative, or unexpected ways that you have seen teams address the production needs of RAG systems? What are the most interesting, unexpected, or challenging lessons that you have learned while working on RAG evaluation? How do you see the future of RAG, context engineering, and data-driven agents changing in the near to medium term?
Contact Info LinkedIn
Parting Question From your perspective, what are the biggest gaps in tooling, technology, or training for AI systems today?
Links Naive Bayes Hands on RAG for Production (affiliate link) RAG == Retrieval Augmented Generation Vectara Hybrid Search TF-IDF == Term-Frequency Inverse Document Frequency BM25 UMBRELA AutoNuggetizer Agentic RAG Malloy
The intro and outro music is from Hitman's Lovesong feat. Paola Graziano by The Freak Fandango Orchestra/CC BY-SA 3.0
More from AI Engineering Podcast
Summary In this episode, I spoke with Jonathan Ellis about how AI agents are changing the economics of software development and forcing teams to rethink where human time is best spent. We explored his argument that, now that code generation is dramatically faster, traditional human-centered code review is becoming a poor use of attention. Instead, Jonathan makes the case for shifting review earlier in the lifecycle toward specifications, RFCs, and architecture decisions, while investing more heavily in automated validation, static analysis, fuzzing, and other quality gates that can help verify whether software does what it is supposed to do regardless of how the code was written. Jonathan also shared lessons from building Brokk AI’s tooling, including Bifrost for fast, agent-friendly static analysis and Mjolnir for coordinating multiple coding agents across harnesses and subscriptions. We discussed the emerging need for better agent-oriented collaboration systems, the risks of blindly outsourcing judgment to AI, and the challenge of training newer engineers in a world where agents can do much of the implementation work. Overall, this conversation examines what remains uniquely valuable for humans as agents take over more of the mechanics of software delivery: setting goals, making product and design decisions, building better feedback loops, and deciding what is worth building in the first place.
Summary In this episode Nikunj Bajaj, co-founder and CEO of TrueFoundry, talks about the challenge of building reliable agents on top of inherently variable foundation models. He explores the idea of the agent harness as everything around the model engine: memory and context management, tool and MCP integration, sandboxed code execution, permissions, observability, and guardrails. Nikunj explained how TrueFoundry approaches enterprise AI as a centralized control plane for token traffic, while TrueForge provides an open source, vendor-neutral harness for building agents without locking teams into a proprietary stack. He talks about the complexity of model routing across both the harness and gateway layers, the tradeoffs between unified APIs and provider-specific capabilities, and the architectural decisions that affect cost, accuracy, and reliability. He highlighted the importance of benchmark-driven validation, durable execution, human-in-the-loop workflows, sandbox security, and especially agent identity and least-privilege access across chains of agents. Overall, this was a deep look at harness engineering as an emerging discipline, and why organizations need more than isolated point solutions if they want to operate agents safely, efficiently, and at scale. Announcements Hello and welcome to the AI Engineering Podcast, your guide to the fast-moving world of building scalable and maintainable AI systems Your host is Tobias Macey and today I'm interviewing Nikunj Bajaj about the challenges of keeping your agents reliable when the underlying models aren't and how your harness can help
Summary In this episode of the AI Engineering Podcast, Steven Watt, leader of the Office of the CTO at Red Hat, discusses practical paths to achieving AI sovereignty for organizations. He shares his two-decade experience in AI, highlighting how governments are building GPU platforms and protected data hubs to maintain control over AI workloads. Steve emphasizes why self-managed infrastructure is becoming a strategic necessity as companies outgrow cloud costs and require tighter control over models, data, and compliance. The conversation explores the operational substrate for AI sovereignty, including Kubernetes as the scale-out backbone for LLM serving, bridging the gap with PyTorch ecosystems, observability and policy for non-deterministic systems, and emerging security needs such as confidential inference and agentic identity. They also discuss model and hardware optionality (GPUs, CPUs, and new accelerators), the growing demand for energy-efficient inference, and the importance of open models and post-training to create durable differentiation. Steve identifies access to GPUs as the biggest gap hindering sovereign AI adoption today, emphasizing the need for broad access to GPUs for AI workloads to thrive. The conversation also touches on evolving architectures beyond transformers, the interplay between AI and data sovereignty, consolidation pressures from pilot chaos to standardized platforms, and the societal triad of universities, startups, and sovereign infrastructure.
Summary In this episode of the AI Engineering Podcast, Aman Agarwal, creator of OpenLit, discusses the operational foundations required to run LLM-powered applications in production. He highlights common early blind spots teams face, including opaque model behavior, runaway token costs, and brittle prompt management, emphasizing that strong observability and cost tracking must be established before an MVP ships. Aman explains how OpenLit leverages OpenTelemetry for vendor-neutral tracing across models, tools, and data stores, and introduces features such as prompt and secret management with versioning, evaluation workflows (including LLM-as-a-judge), and fleet management for OpenTelemetry collectors. The conversation covers experimentation patterns, strategies to avoid vendor lock-in, and how detailed stepwise traces reshape system design and debugging. Aman also shares recent advancements like a Kubernetes operator for zero-code instrumentation, multi-database configurations for environment isolation, and integrations with platforms such as Grafana and Dash0. They conclude by discussing lessons learned from building in the open, prioritizing reliability, developer experience, and data security, and preview future work on context management and closing the loop from experimentation to prompt/dataset improvements.
Summary In this episode of the AI Engineering Podcast, Carter Huffman, co-founder and CTO of Modulate, discusses the engineering behind low-latency, high-accuracy Voice AI. He explains why voice is a uniquely challenging modality due to its rich non-textual signals like tone, emotion, and context, and how simple speech-to-text-to-speech pipelines can't capture the necessary nuance. Carter introduces Modulate's Ensemble Listening Model (ELM) architecture, which uses dynamic routing and cost-based optimization to achieve scalability and precision in various audio environments. He covera topics such as reliability under distributed systems constraints, watchdogging with periodic model checks, structured long-horizon memory for conversations, and the trade-offs that make ensemble approaches compelling for repeated tasks at scale. Carter also shares insights on how ELMs generalize beyond voice, draws parallels to database query planners and mixture-of-experts, and discusses strategies for observability and evaluation in complex processing pipelines.