An IT Park University student has created an AI-enabled bot for applicants

The bot that answers applicants' questions was presented to Arkady Dobkin, the founder of EPAM Systems, and is currently being integrated into the university's Telegram bot. Learn more about how the student trained the model and optimized the consumption of tokens in the article below.

While studying at the Faculty of Foreign Philology at another university, Roman Kolesov started considering furthering his IT education. According to him, the lockdown of 2021 revealed the possibilities of information technology and remote interaction.

Roman chose IT Park University for two main reasons:

• The convenient hybrid learning format: online lectures and in-person exams;
• One of the founders is EPAM Systems, a world leader in software development.

Among other advantages of IT Park University are:

- Practice-oriented training;

- Internships in foreign offices of EPAM Systems;

- Study lasts three years;

- Teachers of specialized subjects are currently foreign specialists of international companies.

According to the first semester results, Roman was one of the top 10 students of IT Park University.

Chatbot project that answers applicants' questions

The bot advises applicants on admission to IT Park University. You can ask questions in 3 languages: Uzbek, English, and Russian. The use of AI significantly reduces the waiting time for a response and also relieves the selection committee. The expected accuracy of answers is 95%.

ChatGPT is an artificial intelligence chatbot developed by OpenAI. It was launched on November 30, 2022. It attracted attention for its wide range of features: writing code, creating texts, being able to translate, getting accurate answers, and using the context of the dialogue for solutions.

OpenAI is an American company that develops and licenses technologies based on machine learning. One of the founders is Elon Musk.

Background: How the idea was born.

"When the trend for ChatGPT started, I wanted to test the possibilities of artificial intelligence and created a site with embedded AI. With it, I went to the CEO of IT Park University, Bakhodir Ayupov, so that he could assess the idea and give feedback.

During the call in Teams, we discussed the site I created. We decided to show it to the COO, Mikhail Bogatyrev, who is responsible for implementing information technologies at IT Park University. Mikhail came up with several options for using my solution in real-life conditions. He advised me to make the work applicable and vital. Therefore, I decided to create a bot specifically for university applicants." Roman shared.

Machine learning: Looking for optimal solutions

Mikhail advised me to study several technologies and understand which would best solve the problem. At first, I trained the model using fine-tuning, but this technology did not fit.

First, the results were only partially correct. So, having answered, the AI continued to generate information but no longer relative to the essence of the issue. The situation improved when I put a limit on the line.

Second, it was expensive. When training, you must feed the model a certain amount of data. It costs tokens, and each OpenAI token costs a certain amount of dollars. When creating an account in OpenAI, the user receives $5 for testing, and I spent this amount on Fine Tune.

Then, I opened a second account to get another $5 and try Few-Shot Learning, teaching with the help of prompts (tips) based on the FAQ from the university website. At this stage, the FAQ needed to be supplemented. For example, there was no information about Foundation Studies - a preparatory semester that allows you to enter without exams. Mikhail personally supervised the revision of the FAQ, which allowed me to improve the prompts. Now, the chatbot has complete information to support applicants.

As a result, training the model came out cheaper in Few-Shot Learning, and the answers were correct compared to fine-tuning.

By the way, Mikhail provided me with a paid account, and I worked on the project thoroughly without restrictions. I also wanted to buy a paid OpenAI account, which costs $20 per month, but I could not because of the Uzbek VISA. For this, you may need a European or American card.

Choice: ChatGPT 3.5 or ChatGPT 4.0

Choosing between ChatGPT 3.5 and ChatGPT 4.0, I opted for the first version because it is cheaper and responds with the same high quality. It is clear that the more requests, the more money is deducted from the account. We predict a sharp increase in requests before the entrance exams. However, since each request to the chatbot is cheap (up to 2,600 tokens), it will be relatively inexpensive.

I also wanted the bot to respond not with the exact words but diversly during the dialogue. Therefore, I set up such an indicator as temperature. When it is at zero, the bot replies with the exact phrases. But if the temperature is raised, the user will receive an answer to the same request in different wordings. The core idea remains the same, of course.

Language issue: ChatGPT poorly recognizes Uzbek

Initially, I set up the model in such a way that it sent a request to ChatGPT in the language in which it was specified. AI had to recognize it and give an answer in the original language. We have written algorithms for three languages: Uzbek, English, and Russian. However, it turned out that AI does not work well with the Uzbek language. In addition, more tokens are deducted for requests in Uzbek and Russian than in English.

To ease the whole connection, I linked the Google Cloud API. The process follows: the user sends a request, ChatGPT recognizes the language and passes the information to the Google Cloud API. He translates the question into English and returns it to artificial intelligence. ChatGPT provides the answer, and the Google Cloud API translates it into the original language. Now, I'm working on how to eliminate double access to ChatGPT. I will transfer the language recognition function to the Google Cloud API to speed up the processes. In addition, this will provide support not only in Uzbek, Russian, and English but even more because Google recognizes many languages.

Result: A solution that is being implemented

It took about two weeks from the first idea to the deployment. Now, the EPAM developer is rewriting the chatbot in Java to integrate it into the university website.

On June 11, I presented the chatbot to the founder of EPAM Systems, Arkady Dobkin, during his visit to Uzbekistan. Arkady liked the project; he especially noted that the "Freshman" was developing genuine products. Practice-oriented learning is the main principle of IT Park University.

By the way, CEO Bakhodir Ayupov also tested the demo version. He sent several requests in the Uzbek language and specifically wrote with errors. The bot understood everything and answered correctly.

For me, the primary value of the project lies in gaining experience and immersion in the topic of AI. Solutions with embedded artificial intelligence are the most popular and promising. The project sharply pumped my knowledge. For example, I recently participated in the International Summer Camp from EPAM and learned a lot from AI workshops. Mikhail also said that the global Machine Learning team at EPAM was interested in me and that my future internships could be focused on AI and machine learning.

Connect with @itpu_bot Telegram bot for questions about IT Park University's undergraduate program. Get help from an AI-powered assistant.