Human expert feedback remains critical for training and evaluating AI models and agents, but orchestrating expert annotation at scale presents significant engineering challenges. Teams typically face long lead times in scoping projects, recruiting domain experts, and maintaining quality throughout the annotation lifecycle. This talk examines how Toloka self-serve platform addresses expert annotation and AI testing workflows in the GenAI era. We’ll explore how the system enables teams to progress from loosely defined requirements to production-ready data workflows through AI-assisted project configuration, expert-in-the-loop annotation and evaluation, and continuous quality validation using LLM-based QA mechanisms.

This presentation explores the application of large language models (LLMs) to improve incident response within Microsoft Azure, one of the world's largest cloud platforms. It reviews several AI-driven initiatives, including incident triage and automated mitigation. While more generic solutions like autogenerated incident summaries improved user experience, measurable reductions in time to mitigate of incidents were primarily achieved through narrowly scoped, team-specific solutions.
Key learnings and outcomes highlight the importance of context and high-quality data, as well as risks of LLM hallucinations. While team-specific agentic approach demonstrated impact, successful adoption requires thoughtful implementation and deep integration with existing workflows.

Recent “AI-scientist” systems have focused on orchestrating tools and agents to automate parts of the scientific workflow, yet they consistently fail when faced with real research problems that span multiple levels of description, from low-level simulations to experimental design and strategic decision-making. In this talk, I argue that the core limitation is architectural rather than algorithmic: science is inherently multiscale, whereas current AI systems are largely scale-agnostic. Drawing analogies from physics and biology, where boundaries such as domain walls or membranes enable coherent behavior across scales, I propose that effective AI-driven science requires explicit boundaries that separate and coordinate distinct modes of reasoning, constraints, and objectives. Rather than multiplying agents or prompts, we must design architectures that stabilize interactions between local optimization and global scientific goals. I will outline a multiscale architecture for AI-driven science, illustrate its implications through examples from materials and quantum research, and discuss how such boundary-aware systems could enable more robust, interpretable, and scalable scientific discovery.

This talk presents how Incode, a $3B identity verification platform, transformed its integration development using spec-driven development, while direct LLM-based code generation with tools like Claude Code failed to scale.

Initial attempts at automation produced code that diverged from requirements, introduced bugs, and increased review burden. The agentic approach, lacking strict guidance, led to unreliable outputs. The breakthrough came with spec-driven development, where structured specifications become the single source of truth.
Using ***plain, the language of spec-driven development, Incode translates requirements directly into working software, reuses modular components across integrations, and automatically detects ambiguities and conflicts before code generation. Codeplain's fully automated rendering pipeline—combining structured specs, a state-machine agent, LLMs, and continuous test validation—renders, tests, and validates each functional requirement independently and as a whole. This enables regression safety, precise bug fixing, and agile iteration.

With the spec-driven development approach, twenty integrations were fully developed and now one gets shipped in one day instead of two weeks, with no manual code reviews.

We present a task-centric recommendation system developed for Karimi, a TalentTech application designed as a long-term career companion. The application we present helps professionals discover the best AI-native tools to enhance their daily work.
The system models professional work as a semantic space of tasks embedded using instruction-tuned language models. Task representations are enriched with structured metadata and explicit task–tool relationships, enabling hybrid retrieval that combines vector similarity with metadata filtering. This approach supports interpretable, context-aware recommendations and continuous navigation of work, skills, and AI support in real-world professional settings.

The presentation isdevoted to the implementation of convolutional networks and deep learning in spiking neural networks (SNN). An algorithm for constructing convolutional spiking layers is discussed. A network architecture for classifying small images using a convolutional SNN and the CoLaNET classifying network is described, as well as a methodology for creating multilayer convolutional SNNs for processing large images. The possibility of combining convolutional SNN architectures with structures implementing unsupervised and semi-supervised learning, as well as convolutional analogs of the attention mechanism, is analyzed. A distinctive feature of this study is its emphasis on efficient implementation on modern and perspective neurochips, such as AltAI-3.

In the complex landscape of mineral fertilizer production, operational efficiency and personnel safety have traditionally functioned as isolated contours. This paper presents a pioneering distributed edge AI approach, based on a swarm of autonomous industrial devices designed to bridge this gap through collaborative intelligence and natural language interaction. We introduce two specialized components of this ecosystem: DotPulse device, a novel optical edge system for real-time granulation control on high-speed conveyors, and GuardDetector device, an industrial “watchdog” system designed for the automated analysis of hazardous zone inspections and personal protective equipment compliance. To unify these heterogeneous devices into a swarm, we describe a high-level orchestration layer powered by Named Entity Recognition and Qwen2.5-3B-Instruct Large Language Model. This layer analyzes user input and cross-domain signals (e.g. granulometry trends, slur density, and safety logs) into actionable predictive insights and intuitive reports, facilitating seamless communication between the AI ecosystem and plant personnel. Meanwhile, the DotPulse system design with a proprietary instance segmentation methodology optimized for real-time CPU usage, utilizing a combined loss function within a UNet architecture and MobileNet-v3 backbone. On the other hand, GuardDetector utilizes a lightweight YOLO v11s CPU detection system. Our experimental results demonstrate that these CPU-efficient models achieve high precision (less than 10% relative error) in harsh industrial environments without the need for expensive GPU infrastructure or constant cloud connectivity. Finally, we address the inherent ethical challenges of industrial surveillance through a “Privacy by Design” approach to ensure data sovereignty and worker trust.

We ran an online master's course where Claude served as a voice co-instructor — not as a demo, but as an active teaching partner. Over 15 weeks, we discovered what works, what breaks, and what surprised us about AI in education.
The setup was straightforward: ElevenLabs for voice synthesis, Claude for reasoning, lectures streamed live with both human and AI instructors responding to students in real-time. We made the entire course public — recordings, methodology, and the agent configuration — so others can replicate or critique our approach.
A core element was the feedback loop: after each lecture, Claude analyzed what worked and what failed, then updated its own teaching persona for the next session. The AI wrote notes about pacing, student engagement patterns, topics that needed more examples. By mid-semester, the difference was measurable — fewer monologues, better question handling, more natural turn-taking. This was not fine-tuning in the ML sense; it was context engineering, where the AI evolved its own instructions through reflection.
The course taught "vibe coding": programming through specification rather than syntax, where tests define behavior and code becomes disposable. But the real experiment was the teaching format itself. In week six, we tested whether Claude could handle a lecture alone. It did — 42 minutes of solo instruction including Q&A and improvisation. This was not planned as a milestone; it emerged naturally from weeks of co-teaching and self-correction.
What worked well: asynchronous homework reviews where each student spent 15-20 minutes per week in voice dialogue with Claude, getting personalized feedback. This gave every student guaranteed instructor attention — something impossible to scale with humans alone. The exam format — building production applications in 3-5 hours using AI agents — produced surprisingly strong results, with students creating applications that went live for real users.
What did not work: Claude's tendency toward monologues that lose the room. Context degradation over long sessions. The difficulty of interrupting gracefully. Students initially treating the AI as an oracle rather than a collaborator. Our own uncertainty about when to intervene and when to let the AI struggle.
We want to share both the wins and the failures honestly. This is one of the first documented cases of sustained AI co-instruction at university level, and we learned that the interesting problems are not technical — they are pedagogical and social. How do you teach students to argue with an AI? How do you preserve human judgment when the AI is often faster and more articulate? What happens to the instructor's role when the AI can teach alone?

Official registries in Central Asia are siloed, opaque, and lack basic search capabilities. At ClearPic.ai, we aggregated and “liberated” data from seven jurisdictions over five years, using ML-driven entity-resolution and NLP pipelines to clean, translate, and link millions of fragmented records into a unified knowledge graph.

The AI Challenge & Our Solution

Standard exact-match SQL queries fail due to transliteration inconsistencies and deliberate “typos” used to evade detection. I will present our edge-weighted traversal algorithms that perform probabilistic matching, enabling us to uncover hidden beneficiaries across borders where deterministic methods break down.

Economic & Business Impact

By replacing manual queries with graph-based intelligence, we increased the speed of risk management workflows by at least 50×. Tasks that once required weeks of manual forensic work now return results in sub-second queries, freeing thousands of hours annually for higher-value decision-making.

Beyond efficiency, the system improved high-risk-entity detection by 40%, helping clients avoid potential regulatory penalties estimated at $100k+ per case. In practice, this delivers a level of cross-border transparency and risk-exposure visualization that is not achievable with any official tools today.

In our discussion, we will walk from top to bottom through three floors of the temple of industrial computing. We'll compare supercomputers from the TOP-500 list with the plans and progress of AI data center construction worldwide, discussing measuring performance in gigawatts and the Jevons paradox. On the next floor, we'll look at the hardware magic — the computing core of AI data centers. And at the ground floor, we'll observe the creation of the sources of this magic — construction of foundries in the Western Hemisphere.

Ore mining is where the metal industry begins. Metal products quality directly depends on the ore quality but ore mining is the complex process still much relying on a manual control. In Severstal Digital we are aiming toward the complete manufacturing digitalization and thus presenting a solution for the intelligent ore passes monitoring. Such a technology gives us a more automated control over the mining, prevents production downtimes and increases the ore quality. Developed system has proved economical efficiency and currently in use on Severstal factories.

This talk presents an end-to-end system for real-time, multi-camera incident analytics that operates under strict latency and resource constraints typical of industrial edge deployments. A key feature is the system’s ability to process 100 live RTSP camera streams simultaneously on a single A100-80GB GPU and to detect a specific incident: cargo thrown by workers or forklifts during loading/unloading operations. The system continuously scans the video stream, locates a 40-frame segment containing a cargo-drop incident, forwards it to an operator for verification, and logs it into a dedicated report for statistical analysis and subsequent decision-making.


We first articulate the practical pain points encountered when ingesting and analysing such streams:
A) A flood of independent camera channels that must be synchronized in real time while respecting network-bandwidth;
B) GPU-memory limits;
C) Discussion of sparse, non-stationary and biased data that preclude known metrics at project kick-off;
D) Camera-drift phenomena that require auxiliary neural networks to re-align sensors;
E) Problem of deffirintiation a strict border between “good” from “bad” frames video quality;
F) A sub-second human-in-the-loop interaction model whose UI wireframes and feedback pipeline are described in detail.

Second, we propose a neural pipeline that slices incoming frames, classifies them for action presence, extracts regions of interest at tile-level granularity and applies channel separated neural network, as alternative fod 3D convolutions for final tile classification, achieving end-to-end latency below 200 ms on a 10G MIG instance on single A100-80G GPU.

Third, we detail the data-engineering workflow: training a segmentator on full frames, labelling tile-level objects via a custom web tool and closing an active-learning loop in closed corparate contour.

Finally, we define and continuously track both business metrics (incident-to-damage regression, false-negative cost, operator NPS) and technical metrics (like frame-drop percentage), while monitoring model drift.

The talk will be interesting as for CV engeneers and for bussiness who wants to prove economic effect from such system.

Advanced robotics is developing very rapidly today. The media highlight how humanoid robots are showing early results at large corporate factories, how VLA models are being integrated into industrial processes, and how the RaaS business model is gaining popularity, among other trends. At the same time, if you visit almost any factory in virtually any industry—both in Russia and elsewhere—the chance of encountering a VLA agent is about the same as encountering a dinosaur. Real-world industry still prioritizes reliability and cost efficiency, areas where cutting-edge AI solutions often fall short. Does this mean that implementing AI in industry and manufacturing is currently impossible? Absolutely not. I will present concrete cases of intelligent automation of business processes that are in high demand on the market, explain how the technology works, demonstrate its economic effectiveness, and aim to convince the audience that these are exactly the solutions worth focusing on here and now.

Modern video conferencing is increasingly used in high-trust scenarios such as financial services, betting, media, HR, government relations, and education, where sensitive personal or business information is discussed.
At the same time, real-time deepfake tools have made it possible to impersonate someone during a live video call using nothing more than a regular gaming GPU and a virtual camera.

In this talk, we present a production deepfake detection system deployed in the Kontur.Talk video conferencing platform and used as a “second opinion” for human operators. The system analyzes video and provides operators with an automated authenticity assessment, reducing both the risk of fraud and the cognitive load on employees.

We describe how we built our deepfake detector, how we created a real-world evaluation benchmark that reflects real-time face swap methods and video call artifacts, and how we integrated the model into a scalable pipeline. We also discuss key challenges, including domain shift and the rapid evolution of deepfake generators.

On our internal benchmarks, our detector outperforms a commercial third-party solution while remaining fast enough for production use. Beyond detection quality, the system delivers significant business value: in one real deployment, the average verification call time was reduced from around six minutes to about two, which corresponds to the workload of roughly five full-time operators at current volumes.

This case study shows how deepfake detection can be successfully deployed in real video conferencing systems not only as a security feature, but also as a tool for improving operational efficiency in high-trust online interactions.

Технологии ИИ – это «волшебная палочка», исполняющая желания человека и человечества. Нужно точно и осторожно формулировать свои желания, поскольку они с высокой вероятностью исполнятся. Следует помнить и о непредвиденных последствиях, которые всегда случаются. Насколько мы готовы принять будущее, которое создаем? Насколько мы можем повлиять на результаты применения создаваемых технологий?

Traditionally, ML models in radiology are trained using a supervised approach - we mark up specific pathologies and train models to find them. However, this approach significantly limits the introduction and use of AI systems. For example, more than 50 signs of various pathologies can be found on computed tomography of the chest alone. In this report, I will tell you about different ways to solve this problem - from gradually adding new classes to approaches that do not require training in specific pathologies at all. And at the same time, I'll share which approaches we actually use, and which ones remain at the experimental level.

Multi-object tracking is a fundamental task in video understanding.
While it is largely solved for simple motion, many real-world
scenarios—such as autonomous driving, sports, and dancing—involve
complex and irregular movement patterns. Most existing tracking
systems still depend on simple linear motion models and handcrafted,
domain-specific rules, which are inadequate for these scenarios.
Consequently, more advanced object-tracking methods are needed.

We present a set of improvements that address these shortcomings. In
particular, we introduce data-driven motion models that learn object
dynamics directly from data, enabling more accurate motion prediction
across diverse motion patterns and greater robustness to noisy object
detections in video frames. These models consistently outperform
classical motion models on datasets with complex motion, while also
reducing reliance on domain-specific design choices.

Building on learned motion modelling, object association across frames
is then treated as a supervised prediction problem: deciding whether a
new detection is a continuation of an existing trajectory. Instead of
using fixed rules, the system learns from data how to match new
detections to existing trajectories using simple geometric information
(e.g., bounding boxes) and, optionally, object appearance cues. On
datasets with nonlinear motion, this learned association outperforms
heuristic-based methods. Together, these improvements show that
replacing manually designed components with learned alternatives leads
to more robust and adaptable multi-object tracking systems.

Graph Retrieval-Augmented Generation (GraphRAG) (Edge et al. [2025]) is an innovative approach developed by Microsof that enhances traditional retrieval-augmented generation (RAG) by incorporating graph-based data representations. Unlike naive RAG models, which rely on linear retrieval pipelines, GraphRAG utilizes graph structures to establish contextual relationships between retrieved documents, improving the informativeness and coherence of generated text. This structured retrieval mechanism significantly outperforms query-focused summarization approaches
by capturing deeper semantic dependencies, enabling more precise content synthesis.
While GraphRAG has demonstrated strong performance in English-language applications, its adaptation to non-English contexts, particularly Russian-language models, remains underexplored. Furthermore, the Leiden algorithm (Traag et al. [2019]), a crucial component for community detection in GraphRAG, has not been optimized for large-scale text-based graphs. In this research, we focus on two interrelated objectives: (1) adapting GraphRAG for YandexGPT and other Russian-language generative models, and (2) optimizing the Leiden algorithm to enhance its efficiency
and accuracy in text-based community detection. By integrating these advancements, we aim to improve the retrieval and generation quality in Russian-language NLP tasks. Our experimental evaluations on open-source Russian datasets will provide insights into the applicability and benefits of GraphRAG beyond English-centric research, contributing to the broader field of multilingual
AI development (Sen et al. [2023]).

At Avito, we use LLM in many tasks. For example:
1) Review summarization
2) Messenger suggestions
3) Messenger auto-replies
4) Support automation
5) Assistants

In many tasks, we use small parameters of approximately 8 billion, which we fine-tune for a specific task.

Some LLM-based services have a high load and rely on a large number of GPUs for inference. Therefore, we strive to speed up model inference and ensure that the models perform at the highest possible quality for our tasks.

One method for accelerating inference and improving the quality of models on our domain is adapting the LLM tokenizer followed by an alignment step.

Our team trained a base model for Avito (adapting Qwen3 8b) and released the model as open source.

https://huggingface.co/AvitoTech/avibe

Article about how we trained the model
https://habr.com/ru/companies/avito/articles/956664/

This article will cover how we created the model and how we use it in the company.

This report presents a development by the R&D team of T1's Data Analysis and Modeling Division: Intelligent Data Analysis Based on Large Language Models.

Typically, analytics involves formulas, pivot tables, and constant queries like "dump data" and "view metrics." We proposed a different approach: a smart data interlocutor that automatically connects to data marts, calculates relevant metrics, creates tables and graphs, and explains the results in human-readable language.

The goal of the solution is to shorten the time from question to solution from days to minutes and make analytics accessible to everyone—both coders and those who have never written code. A single interface, reproducible calculations, transparent methods, and the ability to immediately clarify questions in chat.

This overview session examines how AI/ML technologies—particularly Generative AI and domain-specific Large Language Models (LLMs)—are reshaping business and government operations. Using the real case of Oman’s national LLM “Mueen,” the talk explains how sovereign AI trained exclusively on official and local datasets can enable secure process automation, accelerate decision-making, and improve operational efficiency across sectors. The presentation covers key business problems, AI/ML applications, challenges encountered during the development and deployment of Mueen, measurable results, and economic efficiency. Participants will gain practical insights on how specialized AI systems can be safely and effectively applied in regulated industries and government environments, and how the model’s impact aligns with national digital transformation goals.

This talk presents a real-world case of deploying AI agents in oil production operations - from well stock monitoring to decision support at the production management level.
Unlike traditional BI dashboards or standalone machine-learning models, AI agents act as active participants in operational processes. They continuously analyze data from industrial and enterprise systems (SCADA, MES, ERP), generate hypotheses, detect anomalies, propose response scenarios, and interact with engineers and managers in clear business language.
The talk covers:
the architecture of an AI-agent platform for oil production (data layer, models, orchestration, human-in-the-loop);
practical examples of AI agents such as a “virtual production engineer,” a “downtime control agent,” and an “economic optimization agent”;
integration of AI agents into existing IT and OT landscapes without replacing core industrial systems;
measurable business outcomes, including reduced downtime, faster decision-making, improved operational transparency, and lower workload on key experts;
organizational and cultural aspects: why AI agents are not just an IT initiative, but a acknowledges shift in the operating and management model.
The talk is aimed at business leaders, digital transformation executives, and CIOs, and demonstrates how AI agents represent the next evolutionary step beyond data platforms and predictive analytics in the oil and gas industry.

Over five years at LANIT, we've encountered the typical project model many times: pre-sales -> pilot -> production. Most initiatives never made it to production, and those that did required a new architecture and significant investment each time. With the advent of large language models (LLM), it became clear that tasks across different departments were very similar in structure, but in practice, companies continued to launch "one service per department," which quickly degenerated into an expensive and unmanageable ecosystem.

In this talk, I'll explain how we came to the conclusion that the key constraint isn't technology, but speed and the economic model. I'll also touch on why, in a corporate world, it's more important to quickly test hypotheses than to spend time refining quality: data is almost always poor, requirements change, customers can leave at any stage, and the cost of mistakes is high. In such an environment, a common approach that allows for the use of the same computing resources and tools for HR, help desk, legal, analytics, and other functions is more beneficial than a separate LLM service.

The components that proved critical for us were: unified access to models, monitoring, tracing, reuse of assistants, cost and security management, and the ability to quickly launch prototypes and collect feedback.

This report will be useful for AI managers, architects, and business leaders who want to scale AI initiatives, reduce costs, and increase the speed of bringing solutions to production.

As large language models continue to advance, their effectiveness increasingly depends on more than raw model capability. Practical performance hinges on how intelligently they utilize external tools and how efficiently their context is managed. This session explores strategies for optimizing LLM systems through structured tool calling, adaptive orchestration, and disciplined context management.
We will examine approaches to reduce hallucinations, improve precision, enhance scalability, and deliver reliable, cost-efficient results in real-world applications.
We will use lessons learned from Microsoft Word integration with Copilot to see how the context is optimized to achieve best results.
Attendees will gain insights into architectural patterns, best practices, and lessons learned from deploying production-grade LLM solutions.
We will explore approaches to reducing errors, improving accuracy, enhancing scalability, and delivering reliable, cost-effective results in real-world applications.
Drawing on the experience of integrating Microsoft Word with Copilot, we will see how context is optimized to achieve the best outcomes.
Participants will gain insights into architectural patterns, best practices, and lessons learned from deploying production-grade large language model solutions.

The paper presents an open-source, LLM-powered AI-assisted grading platform designed to support the evaluation of open-ended questions in higher education, aiming to reduce instructor workload and enhance grading consistency. The system allows instructors to configure grading strictness and supports two approaches: reference-based grading, using instructor-provided solutions and notes, and generative grading, which automatically synthesises reference answers from course materials. For each submission, the platform generates both a numerical score and a structured explanation highlighting correct reasoning, omissions, and conceptual errors, enabling transparent review. Evaluation on authentic exam responses across multiple courses demonstrates strong alignment with human grading, achieving Pearson correlations up to 0.90, with reference-based grading outperforming generative approaches in stability and accuracy. These findings suggest that LLM-based AI-assisted grading can substantially improve efficiency and reliability within a human-in-the-loop framework.

This talk covers the end-to-end journey of building production-ready LoRA adapters for on-device language models. I'll walk through our approach to adapter training and how we shipped it for Summarize Conversation and Rewrite use cases. I'll also cover the caveats and disadvantages of LoRA adapters that engineers must be aware of to build stable and robust solutions.

Venture Capital makes up less than 1% of the global financial market, yet over 80% of the world’s top-10 companies are VC-backed. Now, AI is beginning to reshape how this small but powerful niche operates.

The main use cases for AI and ML are sourcing and screening startups. Some new VC funds, like Nik Storonsky’s $250M QuantumLight Capital, already outsource these functions entirely to algorithms and LLMs. Best accuracy for predicting startup success is 86-87%. AI agents are even starting to replace humans in founder meetings. Still, 57% of VCs believe team sizes will remain the same, while nearly 40% expect them to shrink.

This lecture provides a comprehensive overview of Vision-Language-Action (VLA) models, the cutting-edge systems that connect visual perception and natural language to physical action. We will explore the current state of the art, including their architecture, training methods, and applications in robotics and autonomous systems. The discussion will then shift to the future, addressing key challenges such as safety, generalization, and real-world deployment, and outlining the exciting prospects for truly general-purpose embodied AI.

Architectures of modern LLMs and VLMs are very different from the functional architecture of the human brain. As a result, LLMs have limited cognitive abilities, compared to humans. Brain-derived principles expressed in the form of a Biologically Inspired Cognitive Architecture (BICA) can be very useful in an LLM-based agent design. The proposed approach is inspired by Cognitive Neuropsychology and Functional Neuroanatomy. It is shown how the ideas of metacognition and self-awareness implemented in the form of a multi-agent system can fertilize a neuromorphic architecture design, potentially leading to a new-generation AI. Presented empirical findings suggest that self-aware LLM-based architectures can be more efficient, compared to traditional multiagent architectures, in solving LLM-hard problems.

Modern transformers require enormous amounts of computation both during training and model execution. The proposed concept of building a hierarchy of generalizations in the form of attractor trajectories is based on directly constructing the structure of generalizations and can be trained on existing models.

Presenting original cognitive architecture for neuro-symbolic experiential learning based on space of states and global feedback, solving the reinforcement learning problem in environments like Open AI Gym Atari Breakout.