Nalmai

You might have all the knowledge and qualifications, and above all, a genuine passion to teach and help students grow and learn from your knowledge. All of this is necessary but not sufficient to be a good teacher, as when it genuinely comes to teaching, you have different students from different backgrounds, different levels of adaptability, and most importantly, different ways of grasping a concept. It is never easy to make sure that what you are delivering and the way you are delivering are going to be effective for every student. We have developed Nalmai to close this gap and keep the teaching dream alive.

During the lecture, it transcribes the audio in real time using gpt-4o-transcribe-diarize model for speech-to-text. It detects if a student did not understand a certain concept or might have confusion somewhere along the way by by using GPT-5.6 to analyze the student’s language and combining this with confusion keywords, response latency, and incorrect poll answers in a deterministic Confusion Confidence Score.

In such cases, it suggests to the teacher a better way to explain the same concepts. GPT-5.6 generates this suggestion using the current concept, the detected confusion evidence, and an appropriate teaching strategy. Furthermore, it detects if the teacher has implemented the suggestion by comparing the teacher’s subsequent transcript with the recommended strategy, and it generates a poll question for the student using a strict structured output. All of this happens live during class!

Next, it updates the students’ performance live during the lecture and keeps track of each student’s progress. It does this using Bayesian Knowledge Tracing, where each student’s own poll answers and language evidence update their estimated mastery of each concept. By the end of the lecture, the teacher can give customized homework for the students.

It also keeps long-term memory for each student’s performance during different lectures. SQLite stores each student’s pseudonymous ID, concept mastery, observations, and progress across sessions. A Teacher Memory Agent retrieves the relevant history and uses GPT-5.6 to generate structured, memory-informed teaching guidance. This shows which student has mastered which concepts and which students might still be struggling with something, so that by the end of the semester, the teacher can take revision classes or follow certain actions.

We have tested and verified the agent with 109 automated tests, real classroom-language validation against 30,000+ TalkMoves utterance pairs, and an authored CCS backtest. All documented with explicit evidence boundaries so it never overclaims causality.