When AI Meets Reality: Vilnius
Performance, Scale & Production Lessons
Join HomeToGo and Aerospike, in collaboration with AWS User Group Vilnius, for an evening of technical talks exploring the challenges of building, testing, and operating high-performance systems at scale - and how AI is transforming the way engineers investigate, troubleshoot, and optimize production environments.
Hear firsthand how high-throughput PHP workloads are built and operated at scale, explore the challenges of maintaining performance and reliability in demanding production environments, discover how AI agents can analyze complex performance benchmarks to uncover hidden bottlenecks, and learn how AI-assisted troubleshooting helped investigate and resolve a mysterious AWS production deployment issue.
Connect with fellow developers, engineers, architects, and technology enthusiasts from the local
tech community over drinks and networking.
Agenda
17:00 – 17:30 - Welcome & networking
17:30-17:50 - High-Throughput Aerospike for Modern PHP Workloads
Martynas Žaliaduonis, Senior Software Engineer, HomeToGo
17:50 – 18:10 - Performance Testing Aerospike in the AI Era: From a Mountain of Benchmark Runs to One Root Cause
David Gerchikov, Senior Software Engineer, Aerospike
Performance testing a distributed database produces more data than people can read. Our latest Aerospike study measured server-side backup with an enormous number of runs across a huge range of cluster configurations, record sizes and datasets, from small to multi-terabyte. Every run leaves behind database and kernel-level metrics, logs, latency histograms, device benchmarks and more.
In this talk we show how AI agents now take part in the whole process, from deploying clusters and running tests through our testing framework to analysing the results. Agents check every measurement against the database source code, so a number gets explained, not just reported.
To analyse the data we use dynamic multi-agent workflows: one agent per test configuration, merge agents that look across configurations, and an adversarial verifier that re-checks the headline claims. We'll explain why the way you shard the work decides what you can find. Some of our most important findings were patterns no single agent could see on its own. They appeared only when the results were merged across configurations, and some took us to bottlenecks we hadn't expected.
You'll leave with practical patterns for AI-driven performance analysis: context files over long prompts, citations for every number, never averaging when agents disagree, and writing past mistakes into every agent's instructions.
18:10 – 18:30 - How AI helped restore my faith in common sense
Alexandr Marmuzevich, Data Solution Architect
A true production story from a two-person team running an AI product on AWS with no DevOps of its own. After a routine ECS release, prod served two versions at once for two days: one container task had silently rolled itself back. The talk follows the investigation step by step — a human hypothesis, an AI agent reading CloudWatch logs in fifteen minutes, and a platform team discovering that two failed health checks trigger a silent rollback. Along the way: why "deploy finished, tests green" is not proof of what is running, why our own monitoring missed the signal, and how work splits between an engineer who knows what to suspect and an agent that doesn't get tired of reading logs.
18:30-19:15- Networking
📅 November 24, 2026
🕕 17:00 - 19:30
📍 NFQ Technologies office - Academy Hall, Vilnius