Schedule

Let's figure out what really separates us from general-purpose robots. Let's disassemble a robot and see where each part is reaching its limits, and where progress is happening faster than we think.

We will analyze several typical cases in which our agent and platform use LLM to solve "fuzzy" tasks, and after the talk, the audience will grasp specific approaches based on our production cases.

We will tell and show how the hybrid JIT + AOT code compilation model works under the hood in the Android Runtime and what benefits it brings to real applications. We will compare the Google solution with the Leyden project in the big Java.

I will tell you what we had to modify to run our service locally on Valhalla Early-Access Build, what changes we saw in the metrics, and what conclusions we can draw from this experiment.

How I sketched out a Kotlin Script with a simple agent on Koog. It's not particularly smart, but it's simple and manageable: sometimes it's easier to explain a task with code than with words.

Let's look at the evolution of the JVM through the lens of closing architectural holes. Using the one-nio serialization framework as an example, we will conduct an archaeological dig from Unsafe, MagicAccessorImpl, and synthetic accessors to modern mechanisms such as VarHandles, Nestmates, and Hidden Classes.

We'll take a look at where the FFM API (Project Panama) stands in 2026, drawing on our experience with both using large native libraries like Catboost and more granular — but no less practical — examples, such as our own bindings to libc.

The E2E benchmark, beautiful graphics, impressive RPS and the first place in the table easily create the feeling that the technology is objectively faster than competitors. But what exactly does the E2E benchmark show: the actual stack performance, the quality of the selected scenario, or the ability to prepare the test correctly?

In this talk, we will analyze Java virtual streams not as a feature, but as an engineering solution.

Elasticsearch as a search engine can be debated for a long time. But it turned out to be unexpectedly good for storing metrics: I'll tell you how we launched TSDS under a load of more than a million metrics per second, what problems we encountered and why we eventually had to go inside Elasticsearch and Lucene.

How Java 25 with FFM API learned to manage hardware without a single line of C: we analyze the evolution of the architecture of the open-source Pi4J project: from JNI/JNA to pure Java, the boundaries of applicability, timings and memory — with benchmarks and demos on real sensors.

At the JVM Language Summit 2025, Brian Goetz talked about plans to add type classes to the Java language. The talk is based on Brian's story, as there is very little other information on this topic, but it is not just a retelling.

Why Context Engineering is becoming a new engineering discipline, how AI is changing the entire development lifecycle (AI PDLC), and what expertise will define a strong engineer in the coming years.

The talk will provide an overview of the theory and apply it to specific cases that I have heard in dozens of presentations on the topic, highlighting the common elements. Come to get a ready-made theoretical foundation and a set of practical templates!

The talk focuses on the transition from static agents to self-evolving systems that can adapt to the developer's preferences.

Using the composite case, we will analyze a situation in which an LLM agent dramatically accelerated the preparation of specifications, but almost did not shorten the total product release period.

In reality, only engineers will survive with AI: those who do not fully surrender their thinking to machines and remain decision-makers rather than interface operators.

I'll tell you how to make agents create UI based on guides.

LLMs are destroying not the labor market, but the metaphysical right of humans to exist. Let's analyze this issue from different perspectives.

The dominant paradigm of LLM development assumes that the growth of solution quality is linearly related to the growth of the power (and cost) of the underlying model.

The talk challenges this assumption by showing that with a well-thought-out engineering architecture — the distribution of tasks among short-lived agents, the accumulation of structured state, and strict symbolic control — comparable-quality results can be achieved at a fraction of the cost of models.