Praktika, a language learning startup, is using OpenAI models to power a multi-agent tutoring system aimed at closing the gap between classroom language study and real-world communication, according to a recent OpenAI blog post.
The company was founded by Adam Turaev, Anton Marin, and Ilya Chernyakov, all of whom experienced the challenges of learning English after immigrating to new countries. While traditional education helped them develop reading and writing skills, speaking confidently in professional and everyday settings remained difficult. Praktika was built to address that disconnect by prioritizing conversational fluency.
The app offers daily, interactive conversations guided by AI tutors. Users include students preparing for exams, professionals improving job-related language skills, and immigrants adapting to new environments. Lessons are designed around practical goals rather than scripted exercises.
Multi-Agent Tutoring Architecture
As described by OpenAI, Praktika moved from a single-model setup to a multi-agent architecture to better replicate how human tutors adjust lessons in real time.
The primary Lesson Agent handles live conversations with learners. Powered by GPT-5.2, it combines lesson objectives, learner goals, and recent dialogue to generate natural, unscripted responses. The system can shift exercises mid-session based on engagement or difficulty.
A Student Progress Agent operates continuously in the background, analyzing fluency, accuracy, vocabulary usage, and repeated mistakes across sessions. Using GPT-5.2, it feeds performance insights back into both live tutoring and longer-term instruction decisions.
Long-term progression is managed by a Learning Planning Agent, which uses GPT-5 Pro to sequence skills and adapt lesson plans based on learner goals and historical performance. This agent determines what to teach next and how to balance efficiency with personalization.
All agents share access to a persistent memory layer that stores learner preferences, goals, and prior errors. Rather than loading context in advance, the system retrieves relevant memory only after the learner finishes speaking, allowing responses to reflect the most recent input.
Speech, Memory, and Model Evolution
Speech recognition plays a central role in Praktika’s system. Language learners often hesitate or restart sentences, and pronunciation can be inconsistent. Praktika uses OpenAI’s transcription technology to better handle accented and fragmented speech, reducing interruptions during conversations.
Praktika’s product evolved alongside improvements in OpenAI’s models. Early versions relied on rule-based natural language processing and early generative models, which limited conversational flexibility. The introduction of GPT-3.5 enabled more natural dialogue, while internal testing later showed GPT-4.1 delivered stronger results across onboarding completion, retention, and conversion metrics.
According to the OpenAI post, Praktika saw a 24% increase in Day-1 retention and doubled revenue after introducing long-term memory capabilities. The company now runs its core architecture on GPT-5.2, with GPT-5.2 Pro supporting supervisory reasoning and GPT-5 mini handling continuous progress tracking.
Praktika currently supports millions of learners across nine languages. The company says it plans to expand its AI tutoring capabilities further, with a focus on helping learners use language confidently in real-world situations.
Startup’s approach reflects a broader push to apply generative AI to education in more specialized and interactive ways. Other startups are targeting different segments of the learning market. In children’s education, for example, a team of former Google employees recently launched Sparkli, a generative AI learning app that turns kids’ questions into interactive multimedia “expeditions,” combining audio, visuals, games, and quizzes for learners aged five to 12 and offering teacher tools for classroom integration.
