The platform
One engine for lessons that are spoken, not typed. Everything a product needs — voice, face, subject, assessment, data, delivery — sits behind interfaces we can swap without rewriting the rest.
Voice pipeline
The student's microphone goes through voice activity detection and endpointing into streaming speech recognition. The transcript goes to the tutor model, and its words stream into speech synthesis before the sentence is finished.
Each stage is a port with an adapter, so a provider change is one module and one line of configuration. Interruption — stop, flush, cancel — is a hard requirement of the engine, not a feature flag.
Two tiers of model
A fast model answers inside the lesson, where every hundred milliseconds counts. A stronger model works offline: it writes the report after the lesson, judges exercises and plans what to repeat.
The tutor's behaviour is measured: prompts run on frozen sets of lesson situations, scores keep a history, and a prompt change shows what it improved and what it broke.
A subject is a module
Persona, curriculum, lesson phases, correction rules, assessment rubric and progress schema are data and configuration, not code paths. Studify English and the prep2go exam share the engine and nothing else.
Avatars
Basic: a 3D character rendered in the browser from visemes, with no per-minute cost. Premium: a photoreal avatar driven by the same audio. The lesson does not change between them, and neither does the voice pipeline.
Assessment
Pronunciation is scored on target phrases: sounds phoneme by phoneme, fluency, completeness, stress and intonation; the course decides which of them a drill has to pass. Exercises on the whiteboard get a two-sided verdict — what was right as well as what was not. Checkpoints inside a course decide when to move on.
Lesson state and memory
A lesson can be paused and resumed within a month without losing its place. Errors are carried into the next lessons on a schedule until they stop; homework survives between sessions.
Delivery to partners
A product lives in its own repository under a delivery contract. The platform builds it from a pinned release, deploys it to its own project with its own database roles and buckets, and gives the partner's team a preview lane.
Admission links, protocol APIs and payment hooks adapt to the partner's checkout rather than the other way round.
Trust and operations
- Consent by scope and erasure across every store, with data in EU regions.
- Structured logs with latency marks on every turn; tracing and cost per lesson per product.
- Variable cost per lesson is known and bounded, so there are no unlimited plans.