The Fifth Elephant 2025 Annual Conference CfP

Real-time machine learning—making low latency predictions on live data streams—is now considered state of the art. In a hyperlocal delivery app, this means allocating drivers to orders in milliseconds using live data like location, traffic, and past performance. In this talk, I’ll share our first-hand experience building and deploying a feature platform to support these real-time ML use cases at scale. Along the way, I’ll clarify what terms like “real-time” and “near real-time” mean in practice.

In most Data Science teams, feature engineering starts with scattered scripts, short-lived pipelines, and model training on local notebooks—making collaboration, reuse, and reproducibility a challenge. I’ll share how adopting a feature platform addressed these pain points by providing a central, reliable way to build, share, and serve features across both training and inference. This allowed data scientists to spend less time on setup and more time on building models.

I’ll dive deep into key architecture questions: How do you version and test features like code? How do you track freshness, delays, and schema changes in streaming pipelines? What’s the right approach to data retention, job orchestration, and latency budgets? How do different windowing strategies for aggregating data impact storage and latency? How to optimize for compute and storage cost?

Finally, I’ll share lessons learned from deploying a feature platform 0→1 to power 1M+ order allocations a day with real-time predictions.

Takeaway

• You’ll leave with a clear mental model of what goes on inside a feature platform and how to evolve your own ML infrastructure — that spans both inference and training
• Navigating the trade-offs in building ML features on such platforms as opposed to hand-rolled data pipelines

Intended audience

Data engineers, ML practitioners, and platform teams looking to standardize and scale their ML feature infrastructure.
Engineering leaders exploring to invest in MLOps or feature platforms will benefit from the insights into costs and complexity.

Bio

Siri is a backend engineer turned AI builder. She believes AI has the potential to reshape how we interact with technology — not by forcing people to adapt to machines, but by building products that naturally fit into everyday life.

At nilenso, she has worked with a hyperlocal logistics company to build event-driven architectures, set up a data science platform, and improve developer testing tooling. Her experience spans building products and platform services that support millions of users per minute.

Outside of work, she’s a generalist who can often be found brewing coffee, logging 5Ks on nearby trails, or exploring new tech gear.

Links

Notes on what will be covered in the talk: link