Miya - helper from the smartphone
The larger the company, the more questions arise for novice employees. Who to contact about equipping the workplace? Where is the salary? How are corporate parties going? Where to eat at lunchtime? Questions are not so hard to tell, but it’s usually impossible to get answers to them in one place, and it’s not convenient to ask constantly about all the little things about your new colleagues until you get comfortable with the team. We decided to help newcomers to adapt to our company and created for this in Telegram a chat bot named Miya - “M.Video and I”. And we did it on the naked enthusiasm and Agile.
Our IT-Directorate has been successfully working for Agile for two years already, during this time the development speed has increased by 20%, and the cost has decreased by one third. And we had an idea - to apply Agile not only in IT, but also in HR, having made a pilot project using this methodology.
We approached the creation of a chat bot as to the creation of a certain product. And from the very beginning they were guided by the Lean Startup methodology. It says that you need to start small and iteratively move towards more. Throwing any ideas, then tried to rank them in order of importance, that is, to determine what is required of this, without which the chat bot would not be a chat bot. The rest was left in the category for the future, some of which, perhaps, are not needed at all. Then we held an installation meeting, created an Agile team, dismantled roles, made a SKRUM board, painted tasks and started work.
In Lean Startup, development iteration is adopted: you release a product with minimal effort on the market and watch how it is accepted. The product is new, conditions could change, so you can not accurately predict the market reaction. And on the basis of this reaction, you draw conclusions - this is a pivot, a certain turn in the development of a product, when you add something, remove, modify, and release a new version of the product. Everything, the first iteration is finished, now you track reaction to the second version.
Now our Mission is in the first iteration. The methodology states that such iterations should be as short and fast as possible in order to adapt the product to the market. But we are not in a hurry to start the second iteration: since the chat bot is made for the internal corporate market, it still has no competition, and the audience will use it anyway. This does not mean that we will not improve Mia - of course we will! - just not as fast as in the case of a commercial product in a competitive market.
To begin with, it was necessary to determine the platform and understand what technologies we would use to implement the chat bot. Firstly, there was an option for self-development in Python, but we showered it as very long. We needed an interface so that HR employees could independently enter information into the chat bot, and when developing themselves, people far from IT would have to work directly with the database, which would lead either to great difficulties or to additional investments and development complexity.
As a result, we chose one of the standard solutions - this is Dialogflow cloud platform .. It was developed by immigrants from Russia, and in 2016 bought Google. Dialogflow allows you to create a chat bot and train it using a cloud neural network. The advantages of the platform include the simplicity of the interface and the possibility of free use.
At first glance, the platform’s capabilities were quite enough for us, although we hadn’t worked with it before ( more about). So simultaneously with the creation of the chat bot, we also mastered the capabilities of Dialogflow, groped the boundaries of its capabilities. They even went to IBM to see what Watson can do. In our opinion, such a serious industrial product as applied to our tasks is nothing different from Dialogflow. It is worth a lot of money, but it does the same thing: it extracts keywords from simple queries, finds context and returns content.
We also tested how comfortable it would be for new employees to communicate with the bot, not in writing, but in voice. It turned out that if you just give the opportunity to speak, then people do not always know what to ask them. Or you need to build complex communication schemes. Even Alice at Yandex, with their experience and resources, can only answer simple questions. Therefore, we stopped only on the written communication and the menu. You can simply write Mie, for example, “the structure of the company,” and she will tell you what departments the company consists of, show some photos. And you can choose from the tree menu any questions and immediately read the ready answers.
Please note that under the text entry field there is a button “To the beginning”, which returns to the very beginning of the menu - no need to scroll up or even restart the bot. Also, when selecting an item, the chat bot duplicates the selection made with text so that the user is not lost.
So far, the bot does not know how to communicate in a dialogue mode, this is a Wishlist for the future. We are also thinking whether the standard Android assistant with a Russian-language speech synthesizer can be connected to Mie so that it can be converted into text and vice versa. This will still teach Mia to perceive live speech. For example, it will be possible to ask: “Where to eat?”, Without any “prefix” in the form of “Ok, Google”. The bot recognizes the word "eat" and will give an answer, perhaps even with a voice.
I mentioned above that we are not in a hurry to launch the second iteration of Mia’s development as soon as possible, because it is the only product of this kind in our “domestic market”. Nevertheless, we will definitely develop the chatbot, and for this we need to receive feedback. Dialogflow is a fairly large platform, and now we use only a small part of its capabilities. Among other things, there is an analytics unit, where you can see how many people used the bot, what topics they were interested in, how fully the chat bot was able to answer the questions. Thus, it is possible to find topics in which Miya “floats” and to train her, as well as to identify new directions for her development.
Big plus Mia - its cross-platform. Although Miyu was originally implemented in Telegram, however, the platform allows you to transfer it to other instant messengers, and even bring the functionality separately to a website.
Our chat bot has been working for only two months, and newbies speak well about this adaptation tool. Miya is not interested in “old men” yet, but we have an idea to equip it with such functionality that the chat bot would accompany the employee not only at the initial stage, but also throughout his work in the company as a “guide” on internal policies and procedures.
Our IT-Directorate has been successfully working for Agile for two years already, during this time the development speed has increased by 20%, and the cost has decreased by one third. And we had an idea - to apply Agile not only in IT, but also in HR, having made a pilot project using this methodology.
We approached the creation of a chat bot as to the creation of a certain product. And from the very beginning they were guided by the Lean Startup methodology. It says that you need to start small and iteratively move towards more. Throwing any ideas, then tried to rank them in order of importance, that is, to determine what is required of this, without which the chat bot would not be a chat bot. The rest was left in the category for the future, some of which, perhaps, are not needed at all. Then we held an installation meeting, created an Agile team, dismantled roles, made a SKRUM board, painted tasks and started work.
Start
In Lean Startup, development iteration is adopted: you release a product with minimal effort on the market and watch how it is accepted. The product is new, conditions could change, so you can not accurately predict the market reaction. And on the basis of this reaction, you draw conclusions - this is a pivot, a certain turn in the development of a product, when you add something, remove, modify, and release a new version of the product. Everything, the first iteration is finished, now you track reaction to the second version.
Now our Mission is in the first iteration. The methodology states that such iterations should be as short and fast as possible in order to adapt the product to the market. But we are not in a hurry to start the second iteration: since the chat bot is made for the internal corporate market, it still has no competition, and the audience will use it anyway. This does not mean that we will not improve Mia - of course we will! - just not as fast as in the case of a commercial product in a competitive market.
Platform
To begin with, it was necessary to determine the platform and understand what technologies we would use to implement the chat bot. Firstly, there was an option for self-development in Python, but we showered it as very long. We needed an interface so that HR employees could independently enter information into the chat bot, and when developing themselves, people far from IT would have to work directly with the database, which would lead either to great difficulties or to additional investments and development complexity.
As a result, we chose one of the standard solutions - this is Dialogflow cloud platform .. It was developed by immigrants from Russia, and in 2016 bought Google. Dialogflow allows you to create a chat bot and train it using a cloud neural network. The advantages of the platform include the simplicity of the interface and the possibility of free use.
At first glance, the platform’s capabilities were quite enough for us, although we hadn’t worked with it before ( more about). So simultaneously with the creation of the chat bot, we also mastered the capabilities of Dialogflow, groped the boundaries of its capabilities. They even went to IBM to see what Watson can do. In our opinion, such a serious industrial product as applied to our tasks is nothing different from Dialogflow. It is worth a lot of money, but it does the same thing: it extracts keywords from simple queries, finds context and returns content.
Say a word
We also tested how comfortable it would be for new employees to communicate with the bot, not in writing, but in voice. It turned out that if you just give the opportunity to speak, then people do not always know what to ask them. Or you need to build complex communication schemes. Even Alice at Yandex, with their experience and resources, can only answer simple questions. Therefore, we stopped only on the written communication and the menu. You can simply write Mie, for example, “the structure of the company,” and she will tell you what departments the company consists of, show some photos. And you can choose from the tree menu any questions and immediately read the ready answers.
Please note that under the text entry field there is a button “To the beginning”, which returns to the very beginning of the menu - no need to scroll up or even restart the bot. Also, when selecting an item, the chat bot duplicates the selection made with text so that the user is not lost.
So far, the bot does not know how to communicate in a dialogue mode, this is a Wishlist for the future. We are also thinking whether the standard Android assistant with a Russian-language speech synthesizer can be connected to Mie so that it can be converted into text and vice versa. This will still teach Mia to perceive live speech. For example, it will be possible to ask: “Where to eat?”, Without any “prefix” in the form of “Ok, Google”. The bot recognizes the word "eat" and will give an answer, perhaps even with a voice.
I mentioned above that we are not in a hurry to launch the second iteration of Mia’s development as soon as possible, because it is the only product of this kind in our “domestic market”. Nevertheless, we will definitely develop the chatbot, and for this we need to receive feedback. Dialogflow is a fairly large platform, and now we use only a small part of its capabilities. Among other things, there is an analytics unit, where you can see how many people used the bot, what topics they were interested in, how fully the chat bot was able to answer the questions. Thus, it is possible to find topics in which Miya “floats” and to train her, as well as to identify new directions for her development.
Big plus Mia - its cross-platform. Although Miyu was originally implemented in Telegram, however, the platform allows you to transfer it to other instant messengers, and even bring the functionality separately to a website.
Our chat bot has been working for only two months, and newbies speak well about this adaptation tool. Miya is not interested in “old men” yet, but we have an idea to equip it with such functionality that the chat bot would accompany the employee not only at the initial stage, but also throughout his work in the company as a “guide” on internal policies and procedures.