Conference program

Methods for working with sensitive data: what kinds of data exist, which methods suit each case, how to pick the right method, and where it can all break. We look at how (and what) not to let providers do with your data. A detailed discussion of how to use LLM in situations where a lot of sensitive data may be involved, and how to avoid leaks, reassure the security team, release the service, and not go broke on hardware. Listeners will get answers to the questions: how to use LLM effectively without training the provider, without breaking the law, without attracting regulators, and while staying friends with the security team; what to do if we want to run neural networks on our own (on-premise)? How to “manipulate” data to remove sensitivity while preserving the essence? How to stitch pipelines for the task, balancing privacy and power? How to create “blinders” for LLM so control is not handed to users?

A new data model makes it possible to build a flexible attribution window, link media impressions with Metrica goals and AppMetrica events, and add new parameters and traffic-quality metrics.

We turn a chain of analytical steps into an automated process: forming queries, analyzing dynamics, and using the results to test product and marketing hypotheses. How a multi-step search agent works — one that forms queries, studies results, returns to search, and collects information from many sources — and how to turn typical competitive research from a manual process into repeatable agentic research.

Customer service quality is a manageable economic parameter, not just a set of operational SLAs. At T-Bank, a large-scale A/B test was run that abandoned a single support standard: different segments got austerity modes (ECO, −30% costs) and premium support (CX+, +70% costs). The experiment proved that investment in quality pays off far beyond the contact center: CX+ mode statistically significantly grows NPV, increases product utilization, and even reduces the share of loan delinquencies. We will break down the evaluation mechanics via LTV proxy metrics and show how to differentiate service so you can safely cut costs and maximize total profit

The risks of modern personal data protection laws (GDPR, 152 FZ, CCPA, PDPA, PDPL, DPDPA) in the context of using external AI models. How legal is it to send personal data to external AI service APIs? How realistic is it to protect against risks through obfuscation of the data sent? We break down the risks, the opportunities, and the reliability of such protection

A step-by-step guide to assembling a Bayesian MMM for “flight” placements and a model diagnostics checklist. A breakdown of working with adstock (the effect of fading memory of advertising) and saturation (diminishing returns on investment)

A system of AI agents around the A/B platform and measuring its reliability via a benchmark based on mutation testing: we inject controlled defects into cases with a known answer and calculate mutation score, false positive rate, and the detection threshold. On this proving ground we compare models and architectures and determine the tolerance limits where agents can already be trusted

A system dynamics simulation model calibrated to actual business metrics, answering the question of what will happen to the metrics if business parameters are varied.

Participants will go through the cycle from labeling user actions to forming product hypotheses and launching an A/B test on a real ecom client of the uxrocket.ru platform. Participants will also perform post-analysis and learn to form hypotheses that lead to growth 3 times more often.

We discuss identifiability, bias and decision-making under uncertainty. Bayesian non-linear econometrics makes it possible to estimate causal relationships, and how to build such a model will be the subject of our conversation: how to construct a model so that it is not only interpretable but also identifiable, how to parameterize it so that it doesn't fall apart during estimation and the bias is overcome, how to make decisions under uncertainty, and how stochastic optimization theory can help us.

A working bundle of open-source tools for agentic coding. The OpenSpec and OpenSpec+ development methodology. Turning chaotic generation into a predictable and controlled engineering process.

In our case we cannot split sellers into groups A and B: that would mean giving different sellers different promotion conditions. So a classic A/B test is impossible. The solution is synthetic control: it makes it possible to measure the effect of changes without directly splitting sellers into test and control groups

Banning ChatGPT is easier than learning to work with it safely. Positive Technologies' chief infosec evangelist Alexey Lukatsky explains how Russian companies build secure workflows with Western LLMs. In the lecture we will look at data anonymization methods, the architecture of AI firewalls and, most importantly, how a product analyst can learn to negotiate with the cybersecurity department so that the favorite phrase "infosec won't approve this" becomes the start of a productive dialogue rather than a cross over AI adoption.

When utilization is at its peak, you cannot show the business multiple-fold growth, nor give partners the required volumes at the expense of free traffic. We share our utilization calculation methodology, approaches to media planning and ways to cool down traffic.

Moving work from manual handovers and repetitive development into configuration, generators and platform tooling. This shortens the path from describing a source to data appearing in the platform and lets us scale the onboarding of typical sources without a proportional growth in manual work.

How a deep analysis of an A/B test audience helps find unexpected experiment effects, make more precise product decisions, and improve the quality and accuracy of recommendations.

We turn AI from a code generator into a managed executor capable of taking a task from a vague business question to a verifiable result.

The metric the business wants to influence and the metric the business can influence with an arbitrary action are usually different metrics. We look at how to build healthy iteration on product hypotheses using a metric graph. Listeners will learn how to extract more information from a set of experiments — what else can be learned about users by looking at the whole set of experiments — and how to measure and decompose the effect of a package of marketing changes on a business metric when no single experiment gives an answer on its own.

Solving the task of eliminating chaos in company data and creating a corporate knowledge graph. The sequence of moving from fixing a specific analytical problem to a shared context for AI tools. The talk will help assess which knowledge needs to be formalized in your company, how to check its quality and which workflow to start implementation with.

The specifics of testing AI copilots and other internal products and company processes, in order to answer precisely the question "did the new tool change employees' work enough to make scaling it worthwhile".

Deploying AI does not protect against leaks if, for the sake of speed, a user copies the whole trade secret into the prompt and grants the app access to the entire disk. The talk examines the architecture for restricting AI services' access to corporate data and two vulnerability mechanisms: providers' open logs and the transmission of hidden instructions into active sessions. The speakers will show how to build a "green corridor" for typical tasks, introduce AI solution passports and test security through simulation of work processes.

A scheme for embedding an AI agent into the request-processing pipeline, with automated quality control and important human checkpoints. We break down the solution's architecture, the logic of automated data quality tests, typical mistakes and the agent's operational limits

We have built a conversational AI assistant into everyday work: an employee asks about sales in plain words, by text or voice, and gets a chart or a numeric answer. Access rights are inherited from BI, the answer is verified by code, and next to each number sits its "passport" (period, filters, source). Using a mistake the model made, I will show how verification works, what we measure in the pilot and where the bot went wrong. And then — the move to AI-native analytics: executives work with data themselves, their questions accumulate in a knowledge base, the business maintains the metrics dictionary, and the analyst becomes the architect of a system that brings hypotheses

A complete interference model (distortion of a result due to overlapping experiments), an original criterion for assessing the influence of one experiment on another, and an internal tool for on-the-fly checking, from the ASL team that builds T-Bank's in-house A/B platform.

Assembling agentic systems that deliver results of stable quality. An engineering breakdown of components and architecture patterns for AI solutions.

Connecting an LLM to a database is not analytics yet: the agent does not know where the data came from, who has access to it and how the metric was calculated. We tell how AI was embedded into every step of the analytical cycle (loading, SQL, orchestration, catalog, dashboards) and what context it needed for that: automatically collected metadata, lineage, permissions at the platform level. We will break down where the agent already replaces manual work, where it makes mistakes and how to verify its answer all the way down to the source table. The experience of DataLens Platform, the largest update to DataLens: an AI-first platform where lakehouse, orchestration and a data catalog were added to the familiar dashboards in a single interface

Review sentiment, contact topic, product category — until recently all of this was labeled by hand, and now it is labeled by language models. Along with that, the status of the numbers computed on such labeling has changed. The dashboard metric and the experiment's target metric are no longer observed — they are predicted. Can you compute statistics on a metric that a model assigned, and what is wrong with that statistics?

Game theory for analysts. In his lecture on game theory, Alexey Savvateev will explain the Nash equilibrium concept in accessible terms and show through a series of examples how the interaction of selfish participants can lead to a collective loss. Braess's paradox, the prisoner's dilemma, the tragedy of the commons — in all these situations equilibrium logic does not guarantee a satisfactory outcome from the standpoint of the global benefit to society. Particular attention will be paid to how changing the rules of the game changes people's behavior, which must be taken into account when designing incentive schemes. No special mathematical background is required, but a willingness to reason will come in handy.

Master class participants will learn about a comprehensive approach to ensuring the basic principles of information security for their own products. The session will cover, among other things, automatic scanning against public vulnerability databases, in-house "mirrors" with already verified packages and other tools. Attendees will get a checklist of critically important points to watch during product development by agents. The master class is accompanied by a large set of cases and examples.

How our brain created knowledge and how the artificial information environment affects it, when AI changes the cultural environment faster than human biology.