Boldizsár Bednárik, Aligned Connections Worldwide

AI coding agents are incredibly powerful, but they act like highly caffeinated, drunk interns. They possess infinite typing speed and patience, but completely lack architectural wisdom, business context, and security awareness.
This talk dismantles the dangerous hype of “vibe coding” and provides a pragmatic, production-ready playbook for integrating autonomous agents into real software environments.
Attendees will learn how to shift their role from writing syntax to acting as system architects and QA testers by enforcing strict AI guardrails.
Key takeaways include:
– The Death of “Vibe Coding”: Why common human mistakes like “context dumping” destroy AI accuracy and lead to spaghetti code and hallucinated dependencies.
– The Architect’s Playbook: Actionable workflows to control agents, including mandatory Markdown planning, Test-Driven Development (TDD) as an absolute boundary, and using persistent context (MCP/Qdrant) to prevent hallucinations.
– Backend Security: Why blind trust on the backend is a critical threat, exploring severe vulnerabilities like Prompt Injection and “Slopsquatting” package attacks.
The future of software development isn’t humans versus AI: it is humans managing AI. This session equips you with the strict workflows necessary to stop babysitting your AI and start architecting secure, scalable software.

Miloš Mandić, Schneider Electric

We deployed agentic coding tools across four business units and 135 engineers, product owners and designers. This talk covers what it took to make it work at scale: standardized agent configurations encoding domain knowledge, security hooks and pre-commit gates, quality guardrails in local development and CI/CD, shared prompt libraries, structured feedback loops across all the structures.

Marina Mijatović, AngryLynx

Modern organizations often operate on multiple legacy systems that evolved independently over many years. These systems typically differ in architecture, data models, programming languages, and infrastructure environments, making modernization projects complex, slow, and risky. Migrating such ecosystems into a unified cloud-based platform requires deep analysis of legacy codebases, schema mapping across incompatible databases, and careful orchestration of data and service migrations.

This talk explores how AI agents can assist engineers in modernizing heterogeneous legacy environments. The presentation is based on a real-world projects involving the migration of several independent legacy systems, each with different architectures and database schemas, into a single cloud-native platform designed to support modern data workflows and scalable analytics.

Despot Marković, EPAM Systems

In this talk, I will present a deep learning approach for estimating geological horizon structures directly from 3D seismic data. Unlike traditional methods that rely on manual interpretation or predefined references, this approach jointly learns both the horizon representation and the deformation (warp) required to reconstruct seismic volumes.

The solution is based on a two-component architecture: a horizon estimator that compresses seismic cubes into a 1D structural representation, and a warping model that reconstructs the cube conditioned on this structure. This formulation draws inspiration from optical flow and reconstruction-based learning, but adapts it to the absence of an explicit reference frame in seismic data .

The training pipeline combines synthetic data with dense supervision and real-world data with sparse annotations, enabling robust learning under limited labeling conditions. Multiple loss components, including reconstruction, perceptual, and structural consistency losses, are used to ensure geologically meaningful outputs.

This talk is intended for engineers and researchers interested in applying deep learning to structured 3D data, as well as those working on reconstruction-based models and real-world ML systems.

Nikola Mekić, XG Dynamics

This talk looks at what happens when AI models are integrated into mobile applications, highlighting common issues such as differences between test data and real usage, performance trade-offs, and practical design decisions.

Elena Menshikova, EPAM Systems

Knowledge transfer is a critical yet challenging aspect of complex projects, often hindered by fragmented communication and inconsistent documentation. This presentation focuses on the transformative role of AI facilitator agents in streamlining and enhancing KT (Knowledge Transfer) sessions. By leveraging advanced natural language processing, AI agents can guide discussions, structure agendas, capture key decisions, and generate actionable summaries in real time. These intelligent facilitators ensure that essential knowledge is systematically recorded and shared, reducing the burden on human experts and minimizing the risk of information loss. Practical examples will illustrate how AI agents can improve collaboration, maintain project continuity, and accelerate onboarding for new team members. Attendees will learn about best practices for integrating AI facilitator agents into their KT processes, the benefits for both technical and non-technical teams, and the future potential of AI-driven knowledge management in dynamic project environments.

Sergey Kozlov, AdTech Holding

At AdTech Holding, launching advertising campaigns from client briefs was a manual process involving multiple teams. We built AutoLaunch — a production pipeline that reads Jira tickets, extracts targeting parameters using LLMs, and creates ready-to-launch campaigns automatically.

In this talk I will cover three things. First, the business case — what the manual process looked like and why we decided to automate it. Second, the architecture — an Airflow-orchestrated pipeline with a two-tier approach for parameter extraction and resolution. Third, and most important — how we approached the decision of where to use LLMs and where to keep deterministic code. This is the core design choice that defines how reliable and maintainable the system is. I will share specific examples of where we chose one over the other and why.

Ekaterina Orlova, CAD Software

We present Smart BI — an AI-powered agent that automates investigation of performance drops using internal statistical data. Teams interact with it via Slack — asking questions about trends or anomalies — while the agent handles the rest: planning the analysis, querying databases, running data transformations, and returning results as charts, tables, or summaries.
The talk covers the agent’s architecture, its planning and tool-selection mechanisms, and lessons learned from production deployment.

Sergei Dulikov, Yandex

Training capable AI agents requires carefully designed environments and evaluation protocols. In this talk, I will review established industry approaches and share practical insights from Yandex. The focus will be on synthetic task generation, environment design trade-offs, and robust methods for scoring agent behavior during rollouts, with an emphasis on aligning metrics with real-world performance.