Build­ing an AI tutor: App that teach­es Eng­lish like a real person

In this arti­cle, we’ll take you behind the scenes of our col­lab­o­ra­tion with Kuba Eng­lish —an AI-dri­ven lan­guage app that brings a per­son­al­ized tutor straight to your pock­et. Built on GPT mod­els, the app doesn’t just sound like Kuba; it cap­tures his laid-back teach­ing style, cre­at­ing an inter­ac­tive expe­ri­ence that feels tru­ly alive. 
Let’s explore the key chal­lenges, solu­tions, and insights that helped shape it.  

Cre­at­ing a per­son­al­ized Eng­lish-learn­ing app goes beyond assem­bling lessons and quizzes — it needs to be engag­ing, adap­tive to user needs, and true to the unique per­son­al­i­ty of its teacher, Kuba from Kuba English. 

We spent a lot of time going over prompts to make sure the con­ver­sa­tions sound­ed nat­ur­al and stayed true to Kuba’s tone, with­out feel­ing too script­ed or gener­ic. Using LangSmith’s play­ground, we test­ed dif­fer­ent con­ver­sa­tion­al sce­nar­ios — struc­tured lessons and casu­al chats with role-play­ing exer­cis­es — refined the prompts, and adjust­ed mod­el para­me­ters to keep the AI tutor’s respons­es authen­tic and engaging. 

Anoth­er chal­lenge was keep­ing the out­puts con­sis­tent and find­ing the right bal­ance between being too strict and too lenient in the app’s feed­back. Dur­ing devel­op­ment, we noticed that the AI some­times cor­rect­ed users too harsh­ly or over­looked impor­tant areas for improve­ment. To address this, we used struc­tured out­puts from GPT, defin­ing a clear response format. 

One of our ear­li­est goals was to cre­ate an AI avatar of Kuba that’s more than just his name and face. Using Eleven­Labs’ advanced voice cloning tech­nol­o­gy, we record­ed and processed Kuba’s speech to cap­ture his nat­ur­al tone and emo­tion­al range. The result is a life­like repro­duc­tion of his voice, with nat­ur­al into­na­tion, adding warmth and per­son­al­i­ty to the AI tutor.

At the heart of the app, GPT mod­els from Ope­nAI act as the ​“brains” of each con­ver­sa­tion. They ana­lyze the con­text and the user’s mes­sages to make sure every response feels nat­ur­al, sup­port­ive, and con­sis­tent with the tutor’s friend­ly teach­ing style. 

As a stu­dent works through the lessons, the AI tutor makes the lan­guage and tasks more chal­leng­ing to keep them engaged. After each les­son, the app reviews the con­ver­sa­tion and sug­gests extra con­tent — like pod­casts or addi­tion­al lessons — based on the student’s mis­takes or areas of poten­tial improvement. 

In addi­tion to struc­tured lessons, users can have freestyle chats to prac­tice real-life con­ver­sa­tions. They can talk to dif­fer­ent char­ac­ters — a strict teacher, a wise philoso­pher, or even a six-year-old child — each infused with Kuba’s sig­na­ture humor and style. Stu­dents can choose from pre­pared top­ics like trav­el, nightlife, sports, and movies, or select their own. 

Most impor­tant­ly, the app encour­ages speak­ing prac­tice: users can talk to the AI through voice input and hear Kuba’s voice in response. This helps tack­le one of the biggest chal­lenges in lan­guage learn­ing — speak­ing fluently.

When stu­dents need help, the app gives them rel­e­vant sug­ges­tions to keep the con­ver­sa­tion going. If some­thing is unclear, it also pro­vides trans­la­tions with exam­ples to show dif­fer­ent meanings. 

We also imple­ment­ed a tar­get­ed feed­back fea­ture. At any point, users can ask the AI tutor to review their lat­est mes­sage. GPT then high­lights mis­takes or prais­es a flaw­less performance. 

After each les­son, the app pro­vides a detailed review, sum­ma­riz­ing the student’s key achieve­ments and recur­ring mis­takes, and offers tips for improve­ment — like ​“Try to include a con­di­tion­al phrase” — all in a friend­ly and sup­port­ive way.

To make feed­back bal­anced and use­ful, we use struc­tured out­puts from GPT. Typ­i­cal­ly, the AI-gen­er­at­ed text varies slight­ly with each response, which is fine for casu­al con­ver­sa­tions. How­ev­er, when we need con­sis­tent out­puts for fur­ther pro­cess­ing, unpre­dictable vari­a­tions can ​“break” the code and become a challenge. 

That’s where struc­tured out­puts come in. By defin­ing a fixed for­mat — such as correct/​praise/​suggest for feed­back — we ensure con­sis­ten­cy across the app and UI ele­ments. This approach not only guar­an­tees reli­a­bil­i­ty but also trans­forms feed­back into action­able insights for future lessons or extra practice. 

Com­bin­ing cor­rec­tions with pos­i­tive encour­age­ment keeps stu­dents moti­vat­ed and con­tin­u­ous­ly improving. 

The Kubo­va Eng­lish app is more than a tes­ta­ment to what AI can do — it shows how tech­nol­o­gy can cre­ate a learn­ing expe­ri­ence that feels per­son­al and tru­ly human. By com­bin­ing GPT’s real-time lan­guage pro­cess­ing, Eleven­Labs’ voice cloning, and care­ful­ly engi­neered con­ver­sa­tions, we’ve made sure the app embod­ies Kuba’s unique teach­ing style. 

This mix of inter­ac­tive dia­logue, tai­lored feed­back, and life­like voice not only guides stu­dents but also encour­ages them to speak con­fi­dent­ly, make mis­takes, and keep improving. 

Look­ing ahead, the pos­si­bil­i­ties for AI in edu­ca­tion — and many oth­er areas — are bound­less. From per­son­al­ized health­care and more empa­thet­ic cus­tomer ser­vice to data-dri­ven solu­tions in finance, AI has the poten­tial to change the way we solve prob­lems and con­nect with others.

At Applift­ing, our goal is sim­ple: to keep push­ing the lim­its of tech­nol­o­gy to cre­ate expe­ri­ences that are mean­ing­ful, inclu­sive, and focused on peo­ple above all.