Gamification in banking. Case “Monitoring a bank’s business loan portfolio using three-dimensional visualization Increase user activity and offer new services

Promotion banking products and the formation of the need for them.

Gamification is one of the most popular this moment in marketing trends. And it was logical for us, as a bank with an active and advanced audience, to support it by offering clients a promotion where the game mechanics are implemented at the proper technological level and are largely personalized.

As a result, many users get their first experience of earning interest on money that is just sitting in the bank. Clients on own experience understand that a savings account is a simple and profitable product. And this is the first step to opening a deposit or account, and to expanding your understanding of banking products in general.

An indirect result is also the user's regular use of online banking, since only there can one see their progress.

Moreover, the result is achieved indirectly with the help of game mechanics, presented in the form of a story about an active lifestyle, which is much more interesting to a certain audience than the opportunity to save and receive interest (this is offered by any bank) or a call to use an online bank.

Gamification is a super topic. It's all about involvement. It's boring to make transactions in a bank, it's boring to use banking products. And people love to compete, people love to compete. It sits inside and very deep. And you can exploit these qualities of people. How to do this in a bank? There are few cases. But my deep conviction is that those who learn to actively engage their clients, including using gamification, can earn a lot of money.

Typical approaches:

1. Loyalty programs with points, miles and cashback as rewards.
2. Interactive contextual training for new features. Welcome scripts.
3. Quests and competitions for clients.
4. Creation of simple useful services with game elements: PFM, accumulation on a goal.
5. Viral promotional games that announce new products in an entertaining way.

In March 2017, we were approached by probably the most difficult client in the entire existence of our service, who represented the interests of one large bank that has a considerable network of branches throughout Russia. We are always happy to welcome any clients, but this time our specialists had to deal with typical banking bureaucracy, which we have never encountered with any of our clients before. However, one cannot help but note the fact that the difficulties that we experienced when working with this bank had their own explanation - such a serious organization could not make changes to its website without considering each of the changes we proposed.

Below we will describe in more detail all the difficulties we encountered when collaborating with IT specialists and bank management and how we more than doubled site traffic in 7 months.

First meeting

Our first acquaintance with the site (and employees of the bank’s PR department) took place in March 2017. At that time, the site had very good traffic from search engines, due to the fact that the bank itself and the domain at that time had existed for about 10 years and all this time the PR specialists of this financial institution were actively working with targeted Internet sites and offline advertising, which gave the site trust and in the future it helped us a lot during promotion.

A direct result of many years of work by the bank’s PR team was that a lot of natural links were placed on the site, which, coupled with the old domain, was able to somewhat smooth out the fact that no one was involved in internal optimization of the site.

Also, the bank had a sufficient number of branches in the regions of Russia, which was reflected in the structure of its website, work with which also had to take some time.

Work on the site

SEO site audit

We started the site audit by checking it for duplicate pages and meta tags. In our practice, we constantly come across sites with duplicate pages, meta tags or content, and the bank’s website is no exception. The engine, which was made quite competently from a security point of view, produced several dozen duplicate pages and even more pages with duplicate Title and Description meta tags (for example, a section and section pages had the same titles and page descriptions, which from the point of view of internal optimization, in general , nonsense).

Having checked the site for duplicates, it’s time to start looking for errors in the content. To the credit of the bank’s copywriters and PR people, almost none were found. “Almost” means that the main errors we found were either poor text formatting, missing titles, or missing meta tags for images. We included all comments on each page in a working file for subsequent discussion of changes with customer representatives.

Also, the site was not noticed, which became justifiably popular in last years micro markup "bread crumbs". For a bank website, which has several sections and subsections, the presence of such markup, although not a mandatory requirement, is nevertheless highly desirable.

The last item on our list of errors was about poor internal linking between pages. Despite the fact that some pages were asking to be linked with each other, in fact, nothing like this was done by the bank’s content managers.

This is where we are done with SEO optimization errors on the website.

Technical audit

With the technical side of site optimization, everything turned out much sadder, because, as we wrote above, the engine, which was quite strong from a security point of view, was very weak from an optimization point of view. Including technical ones.

The first thing we checked was the download speed. She, unfortunately, left much to be desired. No, the site, of course, loaded, but it was quite noticeably inferior in speed to competitors’ sites. Two factors led to this: heavy images on the site and the presence of some modules in the template that slowed down loading.

The second point that our SEO specialists noticed is the absence of robots.txt and sitemap.xml files in the root directory of the site. After some clarification from the bank administrators, it turned out that there were none at all. Well, let’s add one more item to the work file for subsequent conversations with clients.

The third point is to search for all outgoing links from the site pages and analyze each of them. Mostly, broken links were found (both to internal pages of the site and to external sites).

The fourth “jamb” of the client portal was poor-quality adaptive layout. It was noticed that when working with the site on smartphones, some blocks moved out and other errors appeared. Typically, everything was fine on the tablets.

The fifth point was checking the site pages for “weight” (in simpler terms, so that the site does not have pages whose code exceeds 200 kilobytes). Looking ahead, let's say that such pages were found and even promptly (compared to other items) fixed by the bank administrators in the direction of reducing the “weight”.

At these five points, technical problems were exhausted, and we moved on to compiling a semantic core.

Semantic core

The creation of a semantic core for existing pages began with the creation of a site structure, starting with home page and ending with pages of the second level of nesting. Under each page, we grouped several relevant high- and medium-frequency queries, some of which were already posted on them in one form or another, and the rest needed to be placed.

We started the process of collecting key phrases by analyzing the content on our clients’ websites, then switched to collecting semantics from similar pages of competitor sites and finished compiling the core of queries after thoroughly working with keyword collection services. Thus, we had several hundred relevant and competitive keywords on hand that needed to be placed on the bank’s pages intended for individuals and legal entities, and there were still several promising groups of queries for which additional pages could be created.

Internal optimization work

Having conducted two audits and compiled the semantic core of the site, our specialists, in conjunction with IT specialists and bank managers, began work on the site.

The first thing we worked on together was correcting technical errors. The easiest thing to do was increase the site loading speed. We optimized the three-dimensional images in Photoshop, which reduced their final size by almost half, and the modules that slowed down the work of the site were partially removed and partially rewritten by the bank’s IT specialists themselves. As a result, the client website began to equal the loading speed of the websites of the largest Russian banks. We also quickly solved the problem of the lack of a robots.txt file and a sitemap: we sent instructions for search bots in the form of a file to the bank’s IT specialists by mail and on the same day we saw this file on the site. The bank refused to write a separate module for the sitemap, preferring a free solution from one of the online services.

Things got a little more complicated with the removal of outgoing links from the site. Despite the fact that there were few of them, the process dragged on for about a week. We don’t know why this is so. However, while the bank was deleting the links we specified, we ourselves managed to create a high-quality adaptive layout for smartphones, which the bank tested for another week.

Thus, in two weeks we managed to deal with technical problems site and go directly to working on the content.

Work on the content began with the client flatly refusing to create new pages on the site, preferring to leave the old structure. Therefore, we only had to draw up detailed recommendations for the bank’s PR managers, following which they would have to change the content on each page or delete duplicate pages. In fact, all the recommendations boiled down to what keywords and in what quantity should be included in the text itself and a discussion of what the Title and Description meta tags should look like (we wrote above that they were duplicated) and H1-H3 tags.

We followed the same scheme in the case of manual linking of site pages to each other - we simply sent recommendations on which page to place the link on and an anchor with the URL for the link.

This process took about two more weeks, while all approvals for changing the content on the site went through the chain from us to the responsible managers of the bank. By the way, the bank listened to the vast majority of our recommendations and changed the content as we told them.

Commercial factors

Separately from everything else, we analyzed commercial factors on the bank's website. On the plus side, the site already had a built-in function for a call back and chat with a support operator, as well as for each deposit and credit offer had its own calculator. Of the minuses, it was impossible to contact support service operators using popular instant messengers, and the “Contacts” page did not have a built-in map from Yandex or Google with the location of the bank. For the most part, these disadvantages were eliminated after the main work on the site was completed.

Also, at our suggestion, the bank’s employees removed outdated information about this financial institution in Yandex.Directory and Google My Business and added up-to-date information.

Conclusion

The Yandex.Metrica screenshot below shows how much we managed to increase site traffic and for how long.

Initially, the metric counter was installed by site administrators in December 2016 (this is not visible in the screenshot). Then, for 2.5 months, the metric simply calculated statistics, and already from the end of March (as we wrote above), our team began working on the site. In our opinion, the result could have been much better if it were not for the constant coordination of all our actions with bank managers, the work of bank employees to correct errors, coordination of what they did with our employees, and the like. As a result, the process, which could have taken two weeks at most, lasted for a month and a half (if not more). On the other hand, one can also understand the top managers of the bank - they simply do not have the right to allow strangers to work on the site, preferring to trust their IT specialists.

To date, the only result of the work carried out on the site is, as we have already said, only an increase in traffic by 2.3 times. We do not have data on increasing the client base.

The application, which I will actually write about below, is based on the iDVP (Interactive Data Visualization Platform) platform.

So, let's begin!

The bank issues loans to large legal entities– borrowers. The number of large borrowers at one point in time does not exceed 1,000. To the bank management A convenient tool is needed with which one could see (monitor) a holistic picture of the bank’s loan portfolio. At the same time, it should be possible to move from a view of the loan portfolio as a whole to detailed information on each of the borrowers.

1. Management wants to spend a minimum of effort working with the application, interpreting visual information, and analyzing data.
2. Management wants to see the status of the loan portfolio immediately by simply opening the application, without making a single mouse click.
3. Information should be presented “as much as possible” - on one screen, without the need for scrolling. Already on the first screen, the user should see which borrowers are “problematic”, how “problematic” they are and what their share in the portfolio is in quantitative and value terms.
4. Tools for filtering and grouping data should be convenient and intuitive.
5. The application screens must be “beautiful” so that management can use it to “effectively” present their reports to the founders and shareholders.

I already talked about our approach to data analysis tasks in a previous article, if you wish, you can read it.

Main points of this article

1. When examining a customer, we always try to identify the customer’s pain (problem) that can be solved using data analysis. And we create an application that completely solves this problem.
2. To analyze data, we do not use “ordinary” BI reports, but 3D applications. In these applications, visualization of analytical information is performed in the form of 3D objects, combined into thematic interactive scenes (screen forms), interconnected by logical transitions.
3. The solutions we create are based on three principles:

• A visual representation of the customer’s business picture. Already at the first acquaintance with the application, on the first screen, the user should see all the parts of his business that interest him.
• Discovering the causes of the problem. Having selected a problem point, the user should be able to use the drill down function, which allows you to fall deeper into the problem area, and see the causes of problems on the following screen forms.
• Technical aesthetics. The application should cause a wow effect, i.e. should be attractive, intuitive and convenient.

Let me remind you of our application design stages:

1. Setting the task and getting started;
2. Customer survey and work with open sources;
3. Analysis, formation of requirements and documentation;
4. Formation of the final document “Description of the Application”.

Setting the task and getting started

Naturally, the application must satisfy all the specific requirements of the bank management listed above.

Customer survey and work with open sources

The main income of most banks consists of providing loans to companies and the population.

Some banks specialize in lending to the population, others in lending to legal entities.

In this bank, the task of monitoring loans issued to large borrowing companies was especially acute. Borrowing companies belong to various industries industry, in this regard, a portfolio analysis is necessary both by company and by industry sector.

The bank compiles and constantly updates a profile for each borrower, which contains information about the borrower’s reliability, his financial indicators.

Bank analysts also collect information about the movement Money(cash flow) of the borrower and other indicators, build cash flow models. Information is collected from several information systems jar.

Based on the results of the analysis of the collected information, problems with the borrower are identified and the borrower is assigned to one of the 5 “problem zones” used by the bank to group borrowers:

1. Green zone – this zone includes a borrower who has no identified problems that could affect the repayment of the loan;
2. Yellow zone – the borrower has some problems identified;
3. Red zone – the borrower has identified significant problems;
4. Black zone – the borrower with a probability close to 100 percent will not repay the loan;
5. White zone – the borrower’s problem level has not yet been calculated.

Depending on the problem of the borrower, the bank is obliged to place special accounts reserves for possible losses, the amount of which depends on the loan amount and the reliability of the borrower. In this regard, it is necessary to control the size of these reserves and prevent their growth, because A reserve is money that is “dead” for the bank, which it cannot use.

Bank analysts also analyze the borrower’s overdue debt (NPL – Non-performing loans). Based on the results of the analysis, the borrower is assigned to one of 4 NPL zones:

1. Green zone - loan payments by the borrower are not overdue or overdue for up to 4 days;
2. Yellow zone – overdue from 5 to 29 days;
3. Red zone - from 30 to 89 days;
4. Black zone - from 90 days and above.

As a result of considering all the borrower’s indicators, the bank calculates its total rating, which shows how reliable the borrower is.

For each loan, timely payments and compliance with other terms of the loan agreement are monitored.

If the next payment is late, the bank finds out the reasons for the delay and takes action against the borrowing company. These could be fines or tougher terms of the loan agreement.

IN loan agreements“covenants” are also indicated - these are special conditions agreements that prohibit the borrowing company from taking actions that negatively affect the borrower’s ability to repay the loan. Examples of covenants are: the borrower’s obligation to provide financial statements to the bank, closure of accounts in other banks, prohibition on obtaining loans from other banks, provision collateral loan.

Analysis, requirements generation and documentation

The more “bad” loans a bank has in monetary terms, the worse the quality of its loan portfolio. Therefore, bank management needs to immediately see “bad” loans and “bad” borrowers, be able to look at the detailed situation of the problem borrower and make a decision on further actions in relation to it.

We decided that the user-manager's work with the application should ultimately be similar to the game "find problem borrower and find out what his problem is.”

Also, to make the application convenient for bank management, we decided to make not only a desktop version for Windows, but also for Mac OS, iOS and Android. Moreover, the platform on which we develop these applications allows us to do this, as they say, “with one touch.”

Based on the results of the analysis, the following indicators were identified that need to be monitored for each borrower:

1. Debt volume
2. Problem area
3. NPL zone
4. Reserve amount
5. Borrower rating

1. See all borrowers on one screen; it must be remembered that the bank simultaneously services up to 1000 large borrowers;
2. Filter borrowers by debt volume;
3. Filter borrowers by problem areas;
4. Filter borrowers by NPL zones;
5. Filter borrowers by bank branches that issued loans to them;
6. Filter borrowers by industry sector;
7. Filter borrowers by problems identified with them.

As I said above, we attach great importance to the convenience and beauty of the application. Therefore, not only analysts, but also 3D designers and usability specialists are involved in working on application screens.

As a result, we got something like this main screen iDVP.Banks.Credit Processes applications (see picture below).

At first glance, the screen seems quite rich, but at the same time all the information is distributed into zones, which makes it easier to perceive. What zones did you end up with?

In this zone, the bank's borrowers are represented in the form of multi-colored planets (balls). The size of the planet corresponds to the amount of debt on the loan of this borrower. The color of the planet corresponds to the borrower’s problem area. In this case, borrowers of the same color are grouped together so that their share (quantitatively and by amount of debt) in the loan portfolio can be visually assessed. Thus, we solved the problem of “seeing all borrowers on one screen.”

In the same zone there is a filter based on the size of the planets (pay attention to the scale and circle located to the right of the planets). Using this filter you can specify the minimum and maximum size debt for displayed borrowers. You can leave only large borrowers on the screen, for example. The task of “filtering borrowers by debt volume” has been solved.

When you click on any planet, you go to the “Borrower Card” screen (see picture below), which shows detailed information according to indicators characterizing this borrower and his loan.

The task of “transitioning from the general picture of the loan portfolio to a specific borrower” to analyze the situation should be done with a minimum number of clicks” has been solved.

In the initial state of the screen, small planets are not always convenient to click on – they are simply difficult to reach with a mouse or, in the case of touch interfaces, with a finger. To compensate for this difficulty, in the central zone it is possible to zoom in and out (zoom-in and zoom-out) of any part of the planetary system. This is done either using the mouse wheel or, if using a touch screen, using the "pinch" action.

This zone contains a filter based on the color zones of the problem of borrowers. You can click/unsnap the desired/unnecessary problem areas. As a result, only borrowing planets of the colors desired by the user will remain in the central zone. The task of “filtering borrowers by problem zones”, “filtering borrowers by NPL zones” has been solved. An attentive reader will probably ask how we filter borrowers by NPL zones using this tool, because it only filters problem zones. It’s simple: in the upper left part of the screen there is the text “DEBT BALANCE” - this is, in fact, a drop-down list for selecting borrower display modes. The following modes are available for selection:

1. DEBT BALANCE – in this mode, the size of the planets is determined by the size of the debt, and the color of the planets is determined by the problem zone;
2. NPL VOLUME – in this mode, the size of the planets is determined by the size of the overdue debt, and the color of the planets is determined by the NPL zone;
3. RESERVE – in this mode, the size of the planets is determined by the size of the reserve, and the color of the planets is determined by the problem zone;
4. RATING – in this mode, the size of the planets is determined by the rating value, and the color of the planets is determined by the problem zone.

Filter area on the right

This zone contains an accordion filter element, which contains three filters:

1. CA+TB (central office + territorial banks) – using this filter, you can leave on the screen only borrowers whose loans were issued by the central office (bank head office) or territorial banks (branches).
2. INDUSTRIES – allows you to filter borrowers from certain industry sectors.
3. PROBLEMS – this filter allows you to leave on the screen only those borrowers for whom the bank’s analysts have identified certain problems.

The task of “filtering borrowers by bank branches, by industry, by problem” has been solved.

Lower chart area

This zone contains a graph that displays the change in the ratio of problem zones or NPL zones over time. For this purpose, the “stacked line chart” type of chart is used. The colors of the graph correspond to problem areas or NPL areas.

The user has the opportunity to set the slider to any date on the chart, and only those borrowers that the bank had at that time will be displayed in the central zone. The sizes of the planets and their colors will correspond to the amount of debt and the problem area that each borrower had on the selected date.

Below I attach the rest of the application screens: thumbnails and names. And you will have the opportunity to study them and analyze them yourself. If you have any questions about the content, ask them in the comments, I will definitely answer.

Main screen with enabled display of a pie chart of the distribution of color problem zones

Main screen with borrowers filtered by debt volume (on the scale to the left of the planets, the lower display limit is set at 20% of the maximum)

Main screen. Approaching planets (zoom-in)

Borrower card

At the first FinMachine forum held on Friday, the director of the risk modeling department of Sberbank, Maxim Eremenko, and the head of R&D in the field of Data Science, Andrey Chertok, described how the country’s largest bank, using machine learning, among other things, generates claims and finds business partners for its clients.

Andrey Chertok: As part of our participation in one of the bank’s projects, we detect behavioral patterns of bank clients based on their transactions. The first models were associated with the descriptive analysis of transaction behavior. For example, the client did not have car-related purchases - they appeared. This means that he bought a car, and now it is possible, for example, to offer such a client products or services that are useful for car owners.

The next task is to predict certain events, including the fact of purchase itself. In addition to patterns, with the advent of certain MCC codes, it becomes possible to extract enough information from the data interesting stories, including those related to the savings activities of cardholders. That is, we see which of the bank’s clients is saving money and warn certain large purchases. This can greatly enhance the models. The bank can provide a wider range of offers. However, this means that such models must be constantly adapted.

On the slide we see three fairly clear cases: buying a car, renovating an apartment/buying furniture, and treatment costs. It is especially valuable if the client can provide feedback on the products offered to him. Therefore, it is necessary to make models that can take this into account. feedback. In many ways, this is the same principle that underlies the reinforcement learning models that we are now beginning to develop.

Reinforcement learning, which is currently being developed by OpenAI and DeepMind, among others, is a harbinger of AI as they want to see it. The system is not pre-installed with any model of the world, and the system actually knows nothing about it. The system begins to interact with the world, receive feedback, so-called rewards. The system then adjusts its behavior based on how good or bad the rewards are received. In the case of banking products, reward is, for example, how interesting or uninteresting a particular bank offer turns out to be for customers.

By using methods with specific properties that enable reinforcement learning, we can adapt these algorithms in real time. Among the new approaches, it can also be noted that just recently an article by the same DeepMind was published in Nature, where they talk about how elements of a Turing machine were introduced into a neural network. As a result, the neural network was able to have memory, which neural networks lack at this stage.

Case 2. Sales funnel optimization
Andrey Chertok: In this case, we analyze transaction activity, looking for clusters of clients with certain behavior patterns. But in this case we do not connect them with the prediction of any events. For example, we can find clients who fly frequently, travel abroad and frequently convert currencies. Based on this, we make offers to such clients more effectively.

The slides show what patterns we can find and what products we can offer in this case. In general, the story is clear - certain methods associated with clustering are assumed here. Data projection, for example.

Case 3. Optimization of cash circulation
Andrey Chertok:Sberbank has a wide network of ATMs, branches, and a scheme for working with corporate clients. Accordingly, the task arises of predicting tomorrow's demand for cash. The more accurately we make this forecast, the more accurately, let’s say, we can distribute this money. On the one hand, it is important that money does not lie idle in ATMs, but instead we can place it on short-term deposit. On the other hand, we strive to avoid reputational losses - money runs out earlier than planned, and the ATM stops working, and the client remains dissatisfied.

Here we need models that can handle asymmetric errors. The first models are very simple and are based on classical methods of time series analysis associated with their smoothing. Now more accurate approaches are required and machine learning methods are already being actively used. Naturally, such methods must be adaptive, since demand depends both on macroeconomic factors and on parameters such as the location of ATMs in the city and the weather forecast. Combining heterogeneous features gives more significant results than using other machine learning models.

Case 4. Modeling the probability of default for small businesses in real time
Maxim Eremenko: In 2014, everyone was talking about Big Data. In 2015, machine learning became disruptive and on the edge. This year, the main trend was deep learning. Next year, obviously, they will talk about reinforcement learning.

Unlike the three previous trends, reinforcement learning is easy to try on open platforms. Open artificial intelligence, funded by Elon Musk, and the DeepMind platform are trained on computer games using an open API that allows you to get into the game code.

We get a battle between two algorithms. If in the 80-90s we played Pac-Man, now the machine controls it and this algorithm can be modified. DeepMind went a little further along this path and, together with Blizzard, built an algorithm for StarCraft.

Algorithms are trained in such a way as to rationalize them for complete applied problems. In the future, they can be effectively trained on tasks related, for example, to the translation of text information into vectors.

Such tasks are the basis of the Google Word2vec engine, which carries out translation from text information into a vector, search and the entire semantic analysis of the text on which it is based.

But the case itself is a little different. We reviewed the active clients of our portfolio in the B2B and B2C segments, paying particular attention to small businesses that actively exchange payments. And when working with them, we tried to abandon classic credit scoring and analysis financial statements and conducting a qualitative assessment of risks regarding the reputation of the beneficiary, managers and similar parameters. Instead, we began to use some kind of aggregate metric, relying solely on transactions - in essence, doing analytical scoring based on the data available to the bank.

As a result, it turned out that the model based on credit scoring, which ranks clients by probability of default, is practically no different in terms of quantitative accuracy metrics from classic models. Her Gini is almost the same at 60-65%. But if the bank’s own information is enriched with external data, say, from social networks, and used for ranking, then accuracy can be further increased.

In practice, this means that there is no need to waste time assessing risks in terms of classical analysis. You can process the data that is in the system and obtain a statistically equally relevant quality metric.

This model can now only be used to generate a list of pre-approved proposals. If the client says: “ok, I agree,” then the process is more complicated. Over time, if we see that the quality of the stream has remained at the current or more high level, and the model shows more predictive accuracy, then it can be used as an alternative.

Case 5. Natural Language Processing algorithms for analyzing and generating statements of claim
Maxim Eremenko: When using tools for working with text or Natural Language Processing, we were faced with the fact that Sberbank is quite a large number of spends human and time resources on analyzing claims and preparing responses. At the same time, the analysis of most of the plaintiffs’ information, and the statements of claim themselves addressed to Sberbank, can be automated. Do not use the labor of people who enter information about passport data in the operative part statement of claim, but you can extract all this: date of birth, passport data, details and the operative part. At the second stage, to prepare the response to the claims, we proposed using a specific template as an optimization.

Case 6. DefinitionB2B- andB2B-chains
Maxim Eremenko: For active B2B users, you can do more than just an assessment credit risk, but also to select typical patterns of his partner. If we see a company with a similar profile in the portfolio economic activity, and both belong to approximately the same cohort, that is, these are not large investment and small businesses, then we, based on these patterns, select partners and recommend which relationships may be of interest to them.

Case 7. Algorithms for the @SberbankML_Bot chatbot
Maxim Eremenko: Our chatbot is still just learning, but it also does some things that many people already know how to do, for example, forwarding via API to open sources like Wikipedia. If you ask him who Gref or Putin is, he will answer.

We have an internal commitment to our bosses that by the summer of 2017 the bot will be able to conduct a conversation on banking topics, plus will have basic cognitive abilities and will be able to carry out conversations on abstract topics. At the moment, the bot is based in Telegram, but we are already developing our own messenger [where it will be moved].

Case 8. Our algorithms can not only learn themselves, but also write poetry
Maxim Eremenko: This is a more entertaining project. We took a recurrent neural network based on the poems of Pushkin, Lermontov and a little on the Jira chat of the developers themselves, and trained the system to write poetry. At first she did not cope well even with iambic tetrameter, but then even rhyme began to appear. Now he manages to write poetry even about Sberbank.