Conference program

Designing a semantic layer and integrating it with Looker. Covers operational constraints, non-obvious technical challenges and our own mistakes.

Our own event stream instead of AppMetrica — and the minimum requirements for event data so an LLM agent can start answering data questions with no human in the loop.

Approaches and methods for analyzing cohorts, plans, lifecycle stages, behavior, reasons for leaving and more.

We will compare how AI works with a database schema, data, technical metadata and a full semantic layer. We will look at which part of business context AI can recover from the schema and the data itself, and which part it must be given explicitly by the company, without turning the construction of a semantic layer into a multi-year project of describing the entire company.

The talk is devoted to the development of the Whydra library for identifying causal dependencies, optimizing Causal Discovery algorithms and evaluating their accuracy and performance. Practical cases of applying the library in banking are considered: analyzing the impact of credit decisions and marketing campaigns on customer metrics, as well as studying the relationship between the launch of new businesses and their subsequent revenue

How antifraud analytics turns behavioral and content signals into product restrictions, how their effectiveness is measured, and how to evaluate detector quality on historical data.

We cluster tens of millions of baskets and show that instead of clear missions there is a continuum with anchors. We explain how to separate the purpose of an order from basket size and discount, which quality metrics to compute before A/B and why we moved from clusters to the «occasion × product group» matrix

A talk about building a system for evaluating advertising effectiveness based on Marketing Mix Modeling. Using the X5 Digital approach as an example, we look at how to combine advertising and business data, choose a model, evaluate a future campaign and compare budget allocation options in a single service. Data preparation, the model and the user process work together. This approach helps consistently evaluate and compare media plans while preserving the model's constraints and human responsibility for the final decision.

Т-Банк's A/B platform (Hippo) combines expertise in mathematics and work with internal data into a single fault-tolerant pipeline. Hippo includes Multi-CUPED (x2.2 speedup), experiment validity checks and sequential methods for early detection of broken experiments (x3.5 speedup at the cost of 1% power). We will discuss how to find the balance between adoption and quality in practice.

Analyzing agent-loop transcripts is a key skill for working on the quality of AI products. In-house evals are often more important than public leaderboards: a 10% quality gain on a state-of-the-art model can lead to a degradation of user experience. Listeners will learn how to apply LLM to highly non-deterministic tasks (in our case it is a city walk, but the general approaches — your own eval, working with transcripts, fine-tuning tools — can be applied anywhere) and will learn to balance the agentic and deterministic parts of a product.

Delegating analytics to AI does not save you from fundamental errors in statistics. The creator of the VWO stats engine, Ишан Гоэл, breaks down the 5 main traps of data interpretation: from the illusion of averages to the paradoxes of A/B testing and the balance of false positive errors. You will learn which questions to ask any analytical report (generated by a human or a neural network) so as not to take a beautifully formatted but mathematically wrong hypothesis at face value.

Retrospective assessment of the cumulative effect on the product from a series of consecutive experiments. Methods for estimating the cumulative effect and when a counterfactual scenario is necessary.

We will reveal the technical side of building an end-to-end ROAS reporting system that tied ad network spend, clicks and revenue into a single loop, which allowed optimizing the performance budget by 30%. The report became a Single Source of Truth both for manual optimization by buyers and for automated agents that partially manage campaigns via API, which ultimately cut reaction time to drops from days to a couple of hours

I will show how to build marketing AI analytics on top of what you already have running, without breaking integrations. We will assemble calls, deals and web analytics from telephony, Битрикс24, Roistat, PostHog and Метрика into a single loop. We will build a customer path graph in Neo4j: from the first click to the deal. An AI agent will process thousands of touchpoint chains and answer the key business question: «On which segments are you losing money, and where do you actually earn it?».

Dashboards stop being someone's job, while semantics and context have not yet become someone's job. How BI at Авито lived through 2026 in this fog: data content certification, domain AI readiness, skills and knowledge bases, prerequisites for an AI analyst, solved and unsolved problems. No success stories, but with a therapeutic effect