Speak at The Fifth Elephant 2026 Annual Conference

Technologies: Apache Fluss, Apache Flink, Apache Iceberg

Abstract: Most AI systems treat data and computation as separate problems. Data lands in a lakehouse in batches, models run in request/response pipelines with no real sense of time, state, or consistency. Great, right up until you need something to happen in real time. Then features drift, decisions get hard to audit, and reproducing what the model did last Tuesday turns into a small archaeology dig.

This talk flips it: the lakehouse becomes the foundation for real-time AI, not the afterthought. With Apache Flink, Fluss, and Apache Iceberg, data, state, and computation finally live in one place. Events land once in Fluss as a durable, ordered log, get processed continuously, and materialize into Iceberg tables that stay consistent and queryable while they’re still warm. Models read live, versioned state instead of stitching together ad hoc joins, with replay, backfills, and exactly-once included.

And here’s the part everyone overcomplicates: an agent isn’t a new kind of thing. It reads from a log, holds some state, makes a decision, and writes the result back as data. That’s a stateful consumer. A microservice with opinions. It’s almost exactly the same shape as a feature pipeline or an inference loop, so once Fluss is holding the log and the state, agents don’t need a separate orchestration layer babysitting them. They’re just another consumer on the stream.

Takeaways: We’ll walk through the patterns (real-time features, continuous inference, event-driven decisions, and yes, agent-style workflows) on a single streaming lakehouse, and you’ll leave with an AI platform that’s real-time by default: consistent, replayable, and actually safe to ship.

{Which audiences is your session going to beneficial for?}

Presenter: Pedro Mázala

{Add the link to draft slides - PDF/PPT - with comments access}

{Add the link to 2-min elevator pitch video}