Hello, Agent!: The podcast at the intersection of data & agents

In this episode, we talk with Bassam Tabbara, Founder and CEO at Upbound, Founder of Crossplane, and Founder of Modelplane, about what it takes to run AI inference at scale.Bassam walks us through his path from writing BASIC on a ZX Spectrum in Lebanon to building the early automation behind Hotmail at Microsoft, founding Crossplane, and then Modelplane: a new open-source fleet orchestrator built for running any model, on any engine, on any hardware.KEY TAKEAWAYS00:00 Alex introduces Bassam Tabbara, inventor of Rook and Creator of Crossplane, here to discuss his newest release, Modelplane.01:20 Bassam's first exposure to programming: teaching himself BASIC on a ZX Spectrum 48K as a kid in Lebanon.05:35 Control planes exist because human-in-the-loop failure response doesn't scale.07:57 Owning your own intelligence is fundamentally about data control, not just cost.11:05 Modelplane was built because Upbound's enterprise customers are already running inference across multiple Kubernetes clusters.13:45 The AI stack — models, serving engines, infrastructure, and accelerators — is tipping horizontal, with more vendor choice at every layer.22:35 The fleet-wide inference problem was unsolved in open source, so Upbound built Modelplane to close that gap.28:45 Modelplane separates concerns between platform teams and ML teams through two distinct APIs.39:35 AI is a workload, not a new platform, and that reframe changes how enterprises architect around it.48:25 Modelplane is an early, 0.1 release, and Bassam invites the open-source community to contribute.Thanks for listening to "Hello Agent!: The podcast at the intersection of data & agents." If you loved this episode, let us know with a 5-star review! Remember to subscribe so you don't miss an episode. To learn more about Redpanda, visit redpanda.comRESOURCES MENTIONEDModelPlane, Bassam's newly released fleet orchestrator for AI inference:modelplane.aiCrossplane, the open-source control plane framework Bassam created:https://www.crossplane.ioRook, Bassam's earlier open-source storage project:https://rook.io/Upbound website:https://www.upbound.io/Cloud Native Computing Foundation (CNCF), which now hosts Crossplane as a graduated project:https://www.cncf.io/vLLM, one of the open-source serving engines discussed:https://vllm.ai/SGLang, the other serving engine discussed:https://www.sglang.io/#RealTimeData #DataStreaming #Redpanda

Most ML training bugs don't announce themselves. They hide in floating-point rounding, batch distribution drift, and concurrency issues that only emerge across thousands of GPUs. The hard part isn't writing the algorithm — it's building infrastructure disciplined enough to tell the difference between a numerically equivalent optimization and a quietly corrupted model.We speak with Tianshu Yu, member of the technical staff at Liquid AI, about the real work of building reliable ML systems at scale. Tianshu previously led reinforcement learning infrastructure at ByteDance, where he was a core contributor to VERL — one of the most widely used open-source RL training frameworks. He now works on pre-training and architecture design for edge-deployed vision language models.The conversation covers the three-phase structure of reinforcement learning, why checkpoints serve debugging, fault recovery, and model selection simultaneously, and how capability collapse emerges when data distributions concentrate during fine-tuning. We also get into what changes when you design models for edge hardware: the shift from attention to convolution-dominant ops, latency budgets as a first-class architectural constraint, and why a 350M parameter model has meaningfully different engineering requirements than a frontier model.KEY TAKEAWAYS00:00 Introduction.03:10 Reinforcement learning trains models against reward signals rather than ground truth, requiring the system to generate its own training data through rollout.06:00 Capability collapse is a real risk during RL fine-tuning: optimizing on a concentrated data distribution can degrade unrelated model capabilities, including multilingual performance.09:00 Checkpoints serve three distinct purposes in large-scale training: monitoring quality curves, enabling fault recovery after hardware or software failures, and supporting model selection across training steps.13:30 A smooth reward curve is not sufficient evidence of a healthy training run — benchmark evaluations at checkpoint boundaries frequently reveal instabilities invisible in aggregate metrics.17:30 Training system bugs decompose into three categories: data pipeline issues, infrastructure logic errors, and algorithmic design flaws — and disentangling them requires iterative hypothesis testing, not a single diagnostic pass.20:30 ML infrastructure correctness is fundamentally different from software correctness: because everything operates in floating point, a mathematically equivalent optimization can still produce divergent numerics.24:00 Eval benchmarks are a late and noisy signal — 30-minute evaluation cycles and LLM-judge variance make them unsuitable as a primary debugging instrument during active training.32:00 Open-source training frameworks like VERL provide two compounding benefits: a battle-tested abstraction layer that separates algorithm from infrastructure, and community validation that reduces the prior probability of fundamental kernel bugs.39:00 Multimodal architectures unify vision and text by mapping continuous image patches into the same embedding space as discrete text tokens, enabling a shared transformer backbone across modalities.45:00 Edge model architecture is constrained not just by parameter count but by hardware topology: devices without large tensor cores favor convolution-dominant operations over attention, requiring architecture decisions to be made alongside inference profiling.Thanks for listening to “Hello Agent!: The podcast at the intersection of data & agents.” If you loved this episode, let us know with a 5-star review! Remember to subscribe so you don’t miss an episode. To learn more about Redpanda, visit redpanda.comRESOURCES MENTIONEDVERLgithub.com/verl-project/verlRobust LLM Training Infrastructure at ByteDancearxiv.org/abs/2509.16293#RealTimeData #DataStreaming #Redpanda

Durable execution has been solving reliability problems for decades. Now it's the missing primitive for agentic AI systems.We're joined by Jeremy Edberg, C-Suite Advisor* at DBOS, Inc., to explore how durable execution works, why it matters for building production-ready AI agents, and what lessons from Netflix-scale reliability engineering apply directly to the age of agentic coding.*Since recording, Jeremy has become C-Suite Advisor to DBOS, Inc.KEY TAKEAWAYS00:00 Introduction 01:30 DBOS was born from research into replacing the operating system with a database. 04:00 A career spent making things reliable — from Reddit to Netflix — built Jeremy's obsession with automation. 09:00 If you do something twice, automate it; make manual steps look like automation. 12:30 Durable execution lets systems resume from exactly where they last succeeded. 14:00 Replaying saved LLM outputs removes non-determinism from agentic debugging. 16:00 In agentic coding, the bottleneck isn't writing code — it's deployment, testing, and operations. 20:00 DBOS adds roughly 1% overhead to Postgres, making durable execution practical at scale.28:00 Agentic AI created a new class of workloads between instant API calls and long batRESOURCES MENTIONEDDBOS, Inchttps://www.dbos.devTransact libraryhttps://www.dbos.dev/dbos-transactCockroachDBhttps://www.cockroachlabs.comDosuhttps://dosu.dev/ Thanks for listening to “Hello Agent!: The podcast at the intersection of data & agents.” If you loved this episode, let us know with a 5-star review! Remember to subscribe so you don’t miss an episode. To learn more about Redpanda, visit redpanda.com#RealTimeData #DataStreaming #Redpanda