screenshot taken from: https://www.hopsworks.ai/dictionary/feature-store
When
Tuesday, 28th April 2026, 5:30pm to 8:30pm
Where
Zalando Office, Hedwig-Wachenheim-Strasse 7, Berlin, Germany
Hosting Organization
Elastic Berlin User Group
Participation Fee
Free Entrance
Agenda
Host Intro, Talk 1, Talk 2, Talk 3, Talk 4, Panel Discussion, Pizza & Socializing
Topics Covered
Talks by Professionals from Zalando, Elastic and Hopsworks & Check-In Queue Due To Missing Volunteers (Host Intro), Engineering an Agentic Research Harness (Talk 1), Simplifying ML Feature Engineering with a Low-Code, Configuration-Driven Framework (Talk 2), Extracting a Latency-Critical Ranking Service from a Monolith Using Code Agents (Talk 3), Building ML Pipelines in the Age of Coding Agents (Talk 4), GenAI Reshaping Engineering Roles, Expectations & the Shift Toward Product-Oriented Thinking (Panel Discussion)
I’ve learned something today
- „Building Machine Learning Systems with a Feature Store“, written by Jim Dowling and published in November 2025, presents a practical, system-level approach to designing scalable machine learning architectures using feature stores as the central data layer. Jim Dowling is the co-founder and CEO of Hopsworks, a company endeavoring to build an AI Lakehouse platform centered around feature store technology.
- Andrej Karpathy is sometimes half-jokingly referred to as a “high priest of AI” because of his rare combination of deep technical contributions and his ability to clearly teach complex neural network concepts to millions. His roles at OpenAI and Tesla, Inc., along with influential courses like Stanford’s CS231n, have made him a central figure shaping how modern AI is understood and practiced.
- P99 latency is the response time below which 99% of all requests complete, highlighting the slowest 1% that often shape real user experience in any system. In LLM applications as well as other distributed systems, most requests may be fast but occasional delays in steps like data retrieval or processing can create noticeable slow responses, which is exactly what P99 is designed to reveal.
- An agent harness is everything around a model such as tools, memory, execution logic, and infrastructure that turns raw model outputs into useful, goal-directed behavior. In practice, it gives the model capabilities it does not have on its own like maintaining state, executing code, accessing data, and operating reliably over complex multi-step tasks.
- Silent data loss occurs when data is lost, corrupted, or dropped without any errors or alerts, making systems appear to function normally while information is actually missing; this is especially dangerous because problems often only become visible later through incorrect or inconsistent results. For example, during a data transfer, certain fields like
user_notesormetadatamay be silently dropped because they are not defined in the destination schema, leading to incomplete records without any warning. - The BIRD (Big Bench for large-scale Database-grounded Text-to-SQL) benchmark evaluates how well AI systems translate natural language into SQL across large, messy, real-world databases, emphasizing correctness, efficiency, and handling of real data rather than just schema structure. Despite recent progress, results on BIRD indicate that human experts still outperform AI, especially on complex queries requiring deeper reasoning, domain knowledge, and robust handling of noisy data.
- The Zalando Office Tower (BHW Building) is emerging as a prominent venue for a wide range of tech events:
picture taken at venue
Published:
Modified: