Language Boss is not a general-purpose AI product pointed at English. Each part of it answers a specific finding about how second languages are learned — from which factor matters most in a classroom, to how long a word survives after a student first meets it. What follows is the reasoning, with sources.
01
The teacher is the largest variable, so the AI reports to them.
The UK Education Endowment Foundation's 2025 rapid evidence assessment on language learning puts practitioner skill above any curriculum or tool. We took that as a design constraint rather than a compliment. Teachers review every generated exercise before it publishes, add instructions that steer the generator, and annotate one-on-one reports before a parent sees them. The system takes the labour — generating fifteen exercise types, grading open responses, scoring pronunciation at the phoneme level — and leaves the judgement.
EEF, Rapid Evidence Assessment on language learning, 2025
02
Meaning-oriented input beats isolated drilling.
The same assessment finds that approaches with a high involvement load — rich, authentic, engaging input — outperform rote memorisation and decontextualised grammar work. So exercises are generated from the class's own dialogues, passages, and vocabulary rather than from a generic bank, and the later stages ask a student to produce language rather than select it. Image description asks them to describe a scene using target structures; voice chat puts them in an unscripted conversation.
EEF, Rapid Evidence Assessment on language learning, 2025
03
Four stages, from recognition to free creation.
Every concept moves through four cognitive stages, adapted from Bloom's taxonomy for second language acquisition. A drill set walks a student through all four on a single grammar point, which is the difference between recognising a structure on a test and reaching for it in conversation.
Recognition
Can the student identify the concept?
Guided production
Can they use it with support?
Constrained creation
Can they apply it in a structured context?
Free creation
Can they use it independently?
Bloom's taxonomy (1956), adapted for second language acquisition
04
Spacing decides what survives the term.
Vocabulary and grammar return on the SM-2 spacing algorithm. The difference from a standard flashcard app is what feeds the algorithm: instead of a binary right or wrong, the student explains the word aloud and the AI scores accuracy, clarity, and examples, with a penalty for slow recall. The interval that follows reflects how well they know it, not whether they guessed. Previous answers are tracked, so repeating yourself doesn't earn a longer interval.
P. Woźniak, SuperMemo SM-2 algorithm, 1990
05
Pronunciation is measured, not estimated.
Speech is assessed at the phoneme level for accuracy, fluency, and completeness, against American, British, or Australian English as the student chooses. Mouth-shape guides show how a missed sound is formed, dual-locale assessment combines phoneme data from more than one accent model, and every assessment is kept — so a pronunciation trend across a term is a chart, not an impression.
Phoneme-level assessment via Azure Cognitive Services
06
Personalisation means the prompt knows the student.
Every AI call carries a context profile: level, submission history, one-on-one analyses, teacher observations, flashcard retention, class progress, and stated interests. An A1 and a B2 answering the same exercise are graded differently. A student whose teacher noted weak connectors gets exercises that target connectives. It is not adaptive learning as a checkbox; it is the individual attention a class of twenty makes impossible.
Student context engine, applied to generation and grading
Bring it to your own classrooms.
A 30-minute walkthrough, tailored to your program.
