«TAdviser SummIT. Best IT practices in Russia 2025 ": section" Artificial Intelligence "

Generative AI potential exceeds expectations

When you work as a technologist, IT specialist, you are managed by business customers. In the fate of the technical unit, there is rarely a chance to really turn the business around, to prove itself. The technological wave of the new time is not just a trend. This is our chance to be the initiators of change and lead something in the company. We got the previous chance at the moment when companies began to implement mobile applications. The introduction of artificial intelligence is just as cool an opportunity, says Kirill Kibalko, an independent expert who co-moderated this session.

Kirill Kibalko, independent expert

He followed the trends and offered to consider what is happening in the world now. Generative artificial intelligence is increasingly being used as a corporate tool. "I do not know what other industry this year has the same amount of investment as generative AI," the speaker notes. He cites figures from an Accenture report called Making Reinvention Real With GenAI, 2025, which said: 83% of executives surveyed believe GenAI's potential to achieve positive business results exceeds their initial expectations.

Three times more companies than a year earlier are starting to invest not just in artificial intelligence, but in AI-agent architectures that can rethink entire workflows. AI agents based on large language models are becoming the new basis for process automation. Business is now taking a strategic approach to implementing AI. Organizations that successfully scale GenAI solutions are 4.5 times more likely to strategically invest in agent architecture, laying the foundation for continuous reinvention. These are metrics all from the same report from Accenture.

Against the background of rosy news, there are also alarming numbers. So, only 36% of the managers surveyed say that they managed to scale GenAI solutions. Only 13% talk about creating significant value at the enterprise level, not at the level of individual pilots. Why? Kirill Kibalko believes that the reason for this may be the following: only 35% of managers have a roadmap, that is, a plan for how exactly generative intelligence will change their labor force.

I believe that this industry, this technology has not yet been fully disclosed. To reveal it, a more balanced and systematic approach is needed. And AI agents working on flexible, adaptable logic are able to really replace people, the speaker is sure. - Therefore, now it is worth looking at the architecture of the enterprise from a new angle. And when creating a roadmap for the introduction of generative AI, it is important not to think that we will create some other intellectual assistant. It is worth considering how we transform our company when 50 employees suddenly appear on the spot 50 thousand, and for the same money.

It said that only 13% of companies have a roadmap. It turns out that we entered these percentages, because we have an AI implementation card in the company. I will tell you what approach we take to create it. Now we have implemented 37 projects, which means the work of 37 recommendation models that work in chemical production.

Alexander Kotelnikov, Head of Technology Department, Industry 4.0, EuroChem

The results of such work speak for themselves: in two and a half years, the company earned 2.75 billion rubles thanks to this implementation. The target for 2027 is 9 billion. Of the models implemented, 25 have a direct economic effect.

As for the development roadmap, we began this activity in 2022, when we re-assembled our digitalization strategy for production. From the very beginning, it was decided to apply a systematic approach, - the speaker shares his experience. - We looked at the company from different points of view: from the point of view of decision-making, from the point of view of the value chain. This chain was divided into elements, and then we consistently worked out with each of our business units those hypotheses that this chain can improve.

To form portfolios of hypotheses, a problem-oriented analysis is used here for each element of the value chain.

LLMs or agents based on them (AI agents) embedded in enterprise systems can improve decisions by providing meaningful information or developing the best way to act. By using multi-agent platforms (for example, n8n), routine operations are automated, and the AI agent will take into account the context.

Strategic decisions are very difficult, they are difficult to prepare. Agents help collect information because creating a model of the environment of a Eurochem-level company means describing the whole world, says Alexander Kotelnikov. - For this to be possible, we have OSINT modules that parse the entire global news stream. This is 7 million events a year in almost all languages ​ ​ of the world, although we are most interested in Russian, English, Portuguese and Hindi - these are our main markets. Based on this news, we are building an event model of influence on the industry. It is provided to our managers, who rely on it to make a decision.

The key to successful implementation is to find the economic driver of the hypothesis, the speaker believes. - If your solution lies in the plane where the function earns direct money, then you need to find a driver there. In production, for example, these are, first of all, downtime and optimal maintenance of the technological regime. In procurement, this is competitiveness and the problem of untreated stocks.

We go to multi-agent systems. This is an element of our strategy, so we are actively studying this topic and we very much hope that with the help of agents we will be able to significantly accelerate Eurochem, "the speaker emphasizes.

We work in a very tight connection with the business. They're involved in everything. This is what allows us to remain repayable in this story, - sums up Alexander Kotelnikov.

AI will look at your melanoma and advise what to do with plaster

Three years ago, I believed that no one needed mobile applications. There is a mobile layout for sites. Then I changed my mind and now I recommend everyone to make such applications, implement them. The same story with artificial intelligence. I thought it was a toy, and now I've changed my mind and I'll show you where it went.

Alexey Ostroushko, Head of IT, network of clinics "Be healthy"

As for the mobile application, they came a long way and eventually changed two contractors, after which they began to develop the product themselves. Now the monthly audience for the Be Healthy app is 78,000. 77% of clients coming to the network of clinics use this application both in order to make an appointment and in order to consult with artificial intelligence.

We have built artificial intelligence into our free application, which, according to the principle of a traffic light, determines with fairly high accuracy whether this mole urgently needs to be shown to a dermatologist. Analyzed 3,000 skin images over 3 months. At the same time, 12% of those who photographed their skin, then one way or another, end up with a dermatologist. In general, it is worth remembering that melanoma is one of the most aggressive cancers, it should be caught on time, the speaker recalls.

AI communicates with a person in text format, collects a history, suggests what to do right now, and, if necessary, records with a doctor for a convenient day and time. The GPT assistant, through a dialogue with the user, is able to identify symptoms, analyze complaints based on big data and suggest which profile specialist it is better to seek face-to-face advice. Using machine learning algorithms and extensive databases, the GPT assistant quickly and conveniently for the user collects information about symptoms and suggests possible causes of diseases, which greatly simplifies diagnosis and saves time not only for patients, but also for doctors.

This year, the clinic has new services for diagnosing dental conditions, for analyzing cognitive abilities and carrying out work to correct the situation. This is for the mobile app. However, AI also works in the clinic itself. Now there are seven of them. An interesting case occurred in radiology. They began to look at old covid pictures and saw that there are sometimes areas of cancer onset.

The radiologist saw covid, but did not pay attention to other areas, especially since they are small. {{quote 'We received such data and phoned these people from the pictures, invited them for examination, "says Alexey Ostroushko. "We looked into the past with AI. It is good that all CT and MRI images accumulate in the systems. Therefore, we have accumulated knowledge for 3-5 years. Artificial intelligence, applied in the right place and the right tools, generates an interesting one. }}

The experience of developing and using a technical support bot with generative AI based on RAG was presented to the audience by Alexander Simonov, project manager for digital services "Knauf," "Knauf Gips."

The company is the leader of the gypsum industry in Russia and the founder ^[1] the modern practice of finishing work. The Russian business "Knauf" has 20 production enterprises producing a wide range of products for high-tech construction: sheet materials based on gypsum and cement, elements of a metal frame, a full range of products for plastering, including mechanized ones.

Alexander Simonov, Project Manager for Digital Services Knauf, Knauf Gips

We offer not only plaster and putty, but also construction technology. We tell you how to properly assemble a wall, floor, ceiling from our materials. Therefore, with the help of a bot, we solve the problem of proper consulting of the client. Even a product manager cannot always navigate the nuances correctly, "the speaker explains the prerequisites.

The range of the company's products is wide, which cannot be said about the human resource. At the same time, there is a great need for media content. At first, the company attracted a large number of contractors to produce such content. More contractors - more documents and approvals. A long chain of approvals leads to excessive bureaucracy of document management. In general, one clings to another and leads to problems. They began to solve them using artificial intelligence.

They asked all sorts of questions, and then they saw that this was not artificial intelligence, and immediately left. That is, the matter did not even reach the solution of issues, - the speaker admits.

There were other difficulties, for example, maintaining the relevance of scenarios and information time-consuming.

At the end of 2024, a full-fledged generative artificial intelligence was introduced into the KAI chatbot. Now the bot maintains live communication, and does not speak with prepared phrases. The database of questions and answers was collected in all channels for a whole year. The collected user questions and technical support answers gave real vivacity to the bot's answers. The bot answers questions on the company's products, taking into account the technical characteristics and technological recommendations for use. Draws conclusions based on the data obtained. If the bot does not find the exact information it needs, then it formulates a competent logical answer, as from technical support. Gives general recommendations. However, he learns new knowledge quickly, it is enough to upload text data and visuals to the knowledge base. The time saving for correcting bot data is 80%.

Now that the Knauf company has tried artificial intelligence, new plans have begun to form. Here they want to create a combined bot (AI plus scenarios), make a closed bot in Telegram to quickly issue information on products to company employees and dealers. Support users in the telephone channel using a bot. Create an admin panel to work with the AI knowledge base without the help of a contractor in order to upload documents there and control the relevance of knowledge yourself.

How AI helps to lead the state hospitals of the Smolensk region, said Alexandra Antropova, chief specialist of the department "Situation Center of the Governor of the Smolensk Region," SOGAU "Center for Information Technologies." Now all state bodies are obliged to keep their pages in VK and OK. Moreover, the indicators of the publics affect the rating of the governor and other federal ratings.

Alexandra Antropova, chief specialist of the department "Situation Center of the Governor of the Smolensk Region," SOGAU "Center for Information Technologies"

In addition, with the help of such pages, we inform citizens, solve their problems and generally find ourselves closer to them. But in the performance of such duties, we ourselves face problems. Imagine that now in the Smolensk region 2,427 state tablets. That's really a lot. Even a small rural library should have its own state hospital. They sometimes do not have the Internet, but there should be a public, - the speaker outlines the situation.

This is due to the fact that we cannot find 2500 experienced smooshchiki and copywriters. Most often, they simply take someone from the organization's employees. They took me myself, I now also lead the public. And the indicators to which we should strive are constantly growing, - says Alexandra Antropova.

And if 4 posts a week are not a problem to print, then collecting the required number of reactions in the form of not only likes, but also comments is really difficult. Anyone who maintains their own page on the social network knows for sure that sometimes one like is already an achievement. The final chord in this symphony - user reactions should certainly be positive! This is really from the field of fiction.

As a result, we realized that we were not ready to abandon any of these models. They already wanted to buy access everywhere, but then the guys from Just AI turned to us with their Jay Copilot product. And it turned out to be the tool that we need, - the speaker shares the details.

The solution gives access to different models of neural networks in the same interface. Allows you to extract ready-made news from reports. There are different applications here, for example, Copywriter and Illustrator. In addition, the protection of sensitive data with JayGard is pleasantly pleasing.

The convenience and speed of content generation have become a real salvation in the context of a dynamic work schedule. Now I can quickly create posts for social networks and focus on other tasks. Modern algorithms of the system help generate original ideas and formats, which diversifies the content and attracts the attention of subscribers.

After the introduction, Jay Copilot managed to increase the audience engagement rate by 75%. Reaction positivity increased by 58%. The preparation time for one item of content has been reduced by more than 2.5 times. The tool is used for other tasks. For example, he writes the governor's speech based on his own report. Of course, employees initially resisted and did not want to use the product. However, they have been trained and the solution continues to scale.

Neural networks will condemn themselves

Not only government departments have problems that AI can solve, but also non-state pension funds, where employees do not know how to use filters in EDMS and search for the necessary documents for a long time. Alexander Zhitin, head of process analysis, NPF Future, personally proposed to introduce generative artificial intelligence to help them.

Alexander Zhitin, Head of Process Analysis, NPF Future

I initiated this project in August 2024. By October, the terms of reference were formed, the first demo was held. In it, the chatbot gave out completely insane answers. Could create a Python calculator, for example. Of course, we removed this function, and now it answers only questions on documents, - says the speaker and gives the product interface on slides.

The user can always evaluate the response that was received. This button is bolted to each replica from the bot.

In the summer of 2025, the decision was put into commercial operation. The chatbot is based on pre-trained open access LLM and RAG technology. Now the search time for the document has decreased by 66%. "I think this is a cool result," comments Alexander Zhitin. In the future, this chatbot is planned to be replicated on a help desk.

How can a business understand whether AI works well at all? Valentin Kaskov, IT Director, proposes to use cascade testing of neural network models for this purpose. international holding "Special Systems and Technologies." Of course, we have GOST R59898-2021. If you read all the provisions, then this is a good help, it is more frequent to talk to IT specialists in the language of technology - this is how the speaker believes.

Valentin Kaskov, IT Director. international holding "Special Systems and Technologies"

Another approach to assess the quality and performance of LLM was developed by Sber. The method is based on gaining new user experience. The company creates a new AI-powered interface and looks at how much people like it, whether their satisfaction levels increase. It checks the average time of execution of the client path, the share of decisions made without human participation, the share of products and services, the parameters of which are selected for the user, as well as the cost per unit of production. If all this looks uninteresting, then the decision is thrown away.

There is a positive experience of foreign startups that have developed neural networks for such an assessment. But there is a difference in our method that multiagent clusters are used, where each agent is individually assigned a separate neural network, and they interact with each other. They themselves determine the hierarchical structure of the cluster, the layers inside this cluster, and they themselves appoint a neural network-arbiter, which will make the final decision: whether the experimental neural network coped with the task or not, - explains the speaker. - As a result, we get a set of bad and good answers that allow us to determine how well trained the neuron is. Based on such a report, it is also possible to build a further training process.

But in general, the method is applicable, interesting. Use. We use this method at home, - suggests Valentin Kaskov.

Banks guessing on tarot cards

Generative AI in business is used to detect fraud, personalize customer experience, optimize IT, assess risks, and so on. Svetlana Mumrina, head of the development of automated service channels, PSB, not only talked about the fields of application, but also told how artificial intelligence is used in their bank.

Svetlana Mumrina, Head of Automated Service Channel Development, P

]]

Now entrepreneurs, as a rule, conduct all communication with counterparties, partners and their employees in the messenger. Our goal was that customers could solve all their issues without moving to another channel, - explains the prerequisites for creating this sub-product Svetlana Mumrina.

For several years now, the consulting agency Markswebb has identified our solution as the best chatbot for business, the speaker emphasizes.

Katyusha not only makes payments. She also knows how to advise on tariffs, helps in working with a mobile and Internet bank, solves trade acquiring issues, facilitates the entrepreneur's routine tasks with the help of AI assistants and gently sells additional banking products. In chats, Katyusha closes 55% of calls without switching to the operator. In calls, this percentage is less and is 37%. Katyusha made 70 thousand payments to VK in 2024, which is 20% more than a year earlier.

An entrepreneur can delegate his responsibilities to AI assistants by pressing the "GPT for Business" button. "Seller" will generate a description of the product card, responses to reviews, classifies reviews from buyers. "SMM Specialist" will come up with an advertising text and a title for it, create a content plan or logo. "HR Partner" prepares questions for interviews or a description of the vacancy, makes a squeeze from the resume.

In addition, RAGs began to be tested here to help employees to facilitate work on internal documents. Intermediate pilot results with RAG - the proportion of successful responses increased from 40% to 75%, but the proportion of answers misleading decreased to 3%. "Recently, she has completely fallen to zero," says Svetlana Mumrina.

On March 8, nothing worked, I confess honestly, - says the speaker. - But we will not stop, because before the New Year we offered to tell the cards of the tarot, and due to this we attracted 20 new accounts, so it worked.

ChatGPT as a mass trigger

I have analyzed your messages in the CRM system. The main reason for the refused customers is too high costs for GPU resources, says Landavid.

Vladislav Balaev, Head of Practice at the Big Data and Artificial Intelligence Competence Center, LANIT

Another reason is the unpredictability of the results in projects to implement machine learning. "Due to the quality and amount of data at the beginning of the project, we often do not know at all whether the problem is solved or not," comments Vladislav Balaev. Gartner also gives not the most comforting forecast: by the end of 2025, 30% of projects with generative AI will be abandoned at the proof of concept stage.

LANIT decided that something needs to be done about this. The problem of high cost was proposed to be overcome by a kind of centralization. Make unified graphical computing resources that will be available to everyone in the company. For this to work, you also need to have monitoring, tracing and scaling tools. A problem called "long" can be tackled by switching to self-service mode, giving users a tool to create their own AI assistants. To do this, you need to borrow a low-cost pipe editor with LLM, OCR and ASR, as well as a smart RAG search. It is worth adding the ability not only to select models and change industrial pots, but also to share your developments, collect feedback.

LANIT combined all this into a single platform and called it "Landev AI. Silicon helpers. " The platform is included in the Register of domestic software. In addition, the owner company is a resident of Skolkovo, which means simpler taxation.

LANIT also uses its solution internally to facilitate the work of a whole group of its companies, and also offers Silicon Assistants to the wide market. Landavid analyzed the main requests of customers who have contacted here over the past two years. Business needs APIs, smart search services, voice analytics, HR solutions, to classify documents and extract details. And most importantly, there is a need to deploy the product within the enterprise so as not to give your data somewhere out. All these possibilities can be found by referring to "Landev AI. Silicon helpers. "

By the way, Vladislav Balaev himself turned to the platform. She helped him create a presentation for the speech by transcribing and analyzing 64 meeting records, summing up 143 customer records in CRM, interpreting 19 requests in Jira, and so on. As a result, the presentation cost the speaker (or company) 140 rubles.

After ChatGPT thundered on the market in 2022, even those who were far from the IT sphere became interested in artificial intelligence. It turns out that not only the infrastructure part is closed, but also people have become more prepared, in any case, morally, for the onset of the era of artificial intelligence. Gleb Kuzmin, Commercial Director, Giga B2B, proposes to form competencies and apply them in practice in business. In the United States and Europe, 85% of the largest companies are already implementing or piloting initiatives related to generative AI. Entire departments appear that deal only with this type of artificial intelligence.

Gleb Kuzmin, Commercial Director, Giga B2B

The model is further developed in specific knowledge domains. Now we have domains in jurisprudence, finance and medicine, - emphasizes Gleb Kuzmin.

It can also be used in production to optimize production processes by analyzing large amounts of data or to create instructions and manuals for operating equipment. GigaChat is used in logistics, commerce, security - there are many options for operating scenarios.

Somewhere we win, somewhere we lose, and somewhere at the same level, but, in principle, such a table allows us to say that the solution is commensurate with the best open source models that exist in the world, "says Gleb Kuzmin.

GigaB2B is a version of the GigaChat neural network designed for business customers. The model is trained in Russian, under full control of the pretrane process (from data collection to specialized training). It is easily integrated into the customer's business processes, especially when compared with open source models. GigaB2B is a comprehensive solution that begins with an infrastructure (server hardware or Cloud/Public API), gives all the possibilities from using the GigaChat 2.0 integrated neural network and a low-cost platform for creating digital assistants for employees with artificial intelligence, skills and the ability to independently work in client information systems.

If a company sees hype in the market, but does not know how to apply it at home, how to find an economic effect, then it can turn to our expertise, we will close this issue, the speaker suggests.

The model is proprietary, the solution is included in the Register of domestic software and is supplied to the customer's contour. By the way, Russian customers will be able to receive subsidies from the federal budget, which are allocated for the development of AI, reduce income tax payments, as well as participate in the AI Leaders competition and thereby improve their image.

Within the RAG methodology

We looked very carefully at artificial intelligence and how it can be applied to business. We tried predictive models in various business processes, used NLP models to build dialog systems for interacting with users, even made themselves a chatbot that allowed an employee to go on vacation, just taking a seat. And then we thought: if LLMs are so cool and so cool to formulate, then if you put them in the framework of business knowledge, it can be convenient. But for this it was important to build the right system.

Alexander Lutai, Technical Director, LANIT Bee Pi Em

The requirements for a question and answer system were simple, but principled. You need to answer without your own speculation and general considerations, according to the internal knowledge base, regulations, company indicators. You will need to delimit access to data, quickly update information. And all this should cost cheap, and work in a closed circuit.

To create such a system, you can further learn LLM. Only for this you need to collect a data set and expensive hardware. The result will be non-deterministic, and with access rights is also a question. The second approach is to take LLM as a ready-made tool, not to climb into it, not to change anything. The disadvantage here will be a limitation on the size of the context.

Ultimately, they decided to turn to RAG, that is, Retrieval Augmented Generation. "We take the initial knowledge base on which we want to get answers, and beat it into pieces, the so-called chunks," the speaker explains. For each chunk, a vector is calculated and added to a separate store (often it will be a vector database). When a request is received from the user, there are chunks close in meaning. An LLM patch is formed, including both the original request and the chunks, after which the model returns a response.

"Next we went to try this approach for different occasions. First inside the company, "recalls Alexander Lutai. Here it was necessary to unload the back office managers who endlessly answered various questions from employees. A neural assistant based on RAG processed more than 5 thousand such requests from staff, while in 90% of cases people were satisfied with the AI response. If only because the employee now receives an answer to his question within 10 seconds.

The decision was not too expensive. For 2 months, work was carried out for a million, the lease of the iron part amounted to 6 thousand rubles per month. "Of course, it was easier for us, because our internal regulatory documents were used, we know where they lie. The analytics phase was therefore missing. "

Another case is already from the field of industry. The customer was the pipe manufacturer, whose technological process is described in GOSTs and TS. Production engineers need to check these documents. Acceptance of products is also carried out according to GOSTs and instructions. Of course, this is a long time. LANIT Bee Pi Em has integrated the corresponding neuropower into their corporate chat bot. Now it answers, and the response rate is 5-10 seconds. The answer contains a link to the original document so that you can independently check the assistant - is it lying. The accuracy of the answers themselves is 95%.

This case differs from the previous example in details, because I had to work not just with text documents, but with tables and formulas. It is not so easy to save this data in such a way as to use it in the future. Also, when working with other customers, I had to face the fact that the vector database is "flat" and does not display links between documents and chunks, which can also worsen the accuracy of answers.

Despite these factors, the RAG platform successfully solves practical problems. Here it is planned to develop further, for example, to use graph structures of knowledge to build semantic links between chunks, to improve algorithms for extracting information from tables, formulas, diagrams and diagrams.

We also work within the framework of the RAG plus LLM methodology - this allows us to control quality, while we have extensive experience in introducing such a story, so we are experts not only in artificial intelligence, "says Vladislav Belyaev, executive director and co-founder of the AI platform, AutoFAQ.

The platform he introduced is a boxed product, so it easily integrates into a wide variety of communication channels.

Vladislav Belyaev, Executive Director and Co-Founder of the AI Platform, AutoFAQ

I have never seen superexperts differ from each other in estimates by a whole order of magnitude. Something is wrong here. I will tell you about the specifics of the implementation of complex projects and how to calculate economic efficiency, "says Vladislav Belyaev.

We are inclined to history when you count the cost of a specific hardware and protect the ROI for each project. It is both simple and profitable. Less risk of investing in outdated or ineffective technologies, the speaker emphasizes.

Our global advantage is that we position the product as a solution for a business customer, - notes Vladislav Belyaev. - He is invited to close his tasks with minimal resources, a boxed product. At the same time, we are not apologists or fans of some kind of technology. We always offer the customer exactly what will help him, and generative intelligence is not always needed. Conventional AI solves 60-70% of requests in our cases. Generative AI should be used meaningfully.

As for the issues with the equipment, the problem with it is simple - everything is expensive. "For customers, we recommend scale models 22-30B parameters. For example, Gemma 27B or Qwen 3. Smaller models are not recommended due to a sharp drop in quality on difficult tasks. For the operation of embedders (for example, E5) and the models themselves, at least 80 GB of video memory is required (A100 80 GB, H100 80 GB or a bundle 2×RTX 4090), "the speaker advises.

The classic problem we face is that the hardware doesn't pay off. Therefore, it is necessary to select the optimal parameters or work in hybrid mode, when the data is in the loop, and the processing is in the cloud, - advises Vladislav Belyaev.

At the same time, noting that it is important for the customer to get his personal, "sensual" experience, as he called it, on his own data. The company offers to download data at the demodlock stage, test Xplain for a whole two weeks, evaluate the quality of work and, if desired, prepare for the pilot.

Organizations are afraid to face AI errors and lose control

We have a huge number of divisions engaged in the introduction of various kinds of technologies, as well as the creation of products related to AI, "says Andrei Gershun, an expert at the Laboratory of Artificial Intelligence at 1C-Rarus.

Andrey Gershun, expert at the Laboratory of Artificial Intelligence of 1C-Rarus

A typical 1C user, in fact, does not exist. Solutions are used in both large businesses and small companies. However, as for artificial intelligence, common thoughts and concerns can be distinguished. Organizations are afraid to face AI errors and lose control. They fear that the infrastructure is not quite ready. They have difficulty estimating the cost of implementing AI and return on investment. End users are biased in principle. In addition, they do not like the high complexity of the system, the fact that it is a "black box," which is impossible to understand because it is difficult to explain why he made this or that decision. In addition, AI hallucinates.

And for us as an integrator who thinks about working with AI, it is important to take into account such requirements as the supply of solutions to the customer's circuit or in the form of a service. For medium and small businesses, the service is more interesting, and for large ones, of course, the best solution is in the circuit. You cannot send data to public APIs - local or secure model placement is required, explains the speaker.

Artificial intelligence can hear, see, predict, simplify, manage and help. Andrey Gershun demonstrated what solutions they implement for each of these categories. For example, AI sees what kind of food is on a tray in a smart cashier for canteens, predicts the company's marginal income, is ideal for cleaning and normalizing large bases of goods that eliminate chaos, and so on. 1C-Rarus implements different AI solutions that improve the efficiency of business processes.

We ourselves use AI for our own development. There are units that are located abroad, so we have to localize products, and large language models are very suitable for solving this kind of problem, "the speaker shares.

A large number of pre-trained models for video cameras that will not spend server power of the customer were presented by Ivan Shamshurin, project manager of the zool.ai from Programming Store. All these models are enclosed in an intelligent video analytics system zool.ai.

Ivan Shamshurin, Project Manager zool.ai from Programming Store

The solution reviews and analyzes in real time all recordings from CCTV cameras. Improves security and reduces enterprise costs. Zool.ai recognizes faces without the use of hardware terminals, applies modern image processing methods and can be trained for any task. Among its advantages is integration with different types of video surveillance equipment. with access control and control systems, as well as with accounting software on the 1C: Enterprise platform, machines and security systems.

The product produces quality control using computer vision. Generates reports at all stages of production to identify problems early and reduce the proportion of defects. It reveals not only defects, but also the amount of products produced. The system can also control personnel. It monitors the routes of movement of people, analyzes their working hours, monitors whether there are undesirable actions and whether everyone has put on the necessary personal protective equipment.

Modules for security systems draw a valid perimeter and configure bandwidth control. Recognize machine numbers, weapons and abandoned items. Zool.ai can be used successfully in marketing video analytics to increase sales. The system will count visitors, check if the queues at the checkouts are accumulating, draw up a store heat map and analyze empty shelves in the required areas. "We adapt the system to the needs of the enterprise and accompany the project from testing to implementation," the speaker emphasizes. Additional training of personnel will not be required, but the solution will be trained - in order to better correspond to the specific tasks of a particular business.

For example, the client uses not ordinary, orange, but pink helmets. We come and do a beautiful photoset on their subject. Then we train intelligence so that it works with this particular client task, form a model and test it together with the client in an experimental and industrial way, "Ivan Shamshurin concludes his speech.

During the break and at the end of the conference, the participants talked informally, and also had the opportunity to familiarize themselves with the solutions and services of IT suppliers at the stands deployed in the event hall.