One agent that teaches, interviews and verifies - by voice and video.
It talks with the person in front of it and draws on a shared screen while it explains. Most of what it does today is teaching: class 1 phonics to postgraduate cardiology.
Opens an email with the details we need. We reply with a time.
Blending c-a-twith Miss Ria
Pick a rung - the board redraws for that learner.
What a session actually is
The learner talks and the agent answers in the same breath, drawing on a board they both look at while it explains. Nobody types. Nobody watches a video.
Twenty seconds of a class 6 lesson
Tutor“How many small unit squares do you think fit inside it?”
Student“Um... twelve?”
Tutor“Good guess - let’s count it together and check.”
- heard: “twelve”
- expected: 16
- gap: −4
- → guided counting
It didn’t mark her wrong and move on. It dropped to counting one row at a time, watched her keep up, and came back to full pace. That is the difference between a lesson that plays and a lesson that is built around the person in front of it.
Area of a square
Who is on the other side
Same agent, different persona, different simulation - built for the learner in front of it rather than pulled off a shelf.
Most tutors start from zero every session. This one doesn’t.
The miss is not just handled in the moment - it is written down. Next week the tutor already knows where to slow down, and it knows it in chemistry too, not only in maths.
- removed: area: unassessed
- added: area: needs guided counting
- added: carried to: fractions, chemistry
Written after the counting session, read at the start of the next one.
Perceive, recall, ground, compose, speak - every single turn.
Nothing is pre-rendered. When the student said “twelve”, this is the work that produced the next sentence:
- perceiveheard “twelve” · read the board state
- recalllearner model: area — unassessed
- groundretrieved the unit-square method for this grade
- composegenerated a guided count, one row at a time
- speak“Good guess - let’s count it together and check.”
- Realtime speech
- Holds the conversation both ways at once, so a learner can cut in mid-sentence and be heard.
- Grounding swarm
- Small, fine-tuned models curate what gets taught against real source material, so the agent does not improvise facts at a child.
- Orchestrator
- Written in-house. Carries context across a session and between sessions - it is what remembers the learner.
- Generated surface
- The board is composed at runtime from primitives - canvases, plots, widgets. There is no lesson library to run out of.
It runs on our infrastructure, or on yours - the enterprise deployments keep every session inside the customer’s own environment.
It listens. It draws and adapts to your child, it watches and scores your candidate, it reads and verifies your customer.
Most of what it does is teaching. It is not only teaching.
The same loop runs a technical interview, a document check and a customer call - with the visual channel open, or closed. Nothing is swapped out underneath.
Quizzes, mock tests and interviews
It hands over a live SQL or Excel widget and steps back - watching the screen, hearing the reasoning out loud, moving the difficulty with the answers. It ends on a rubric, not a number invented at the end.
- identity locked
- one face in frame
- no second screen
- rubric emitted
Funnel analysis, on a live widget
Verification
The same perception, pointed at a document. It reads each field with a confidence it will show you, checks for a live face and matches it to the card - guided out loud, so nobody is left guessing why an upload failed.
- PAN detected · 3 fields read
- blink detected
- face ↔ card match
- 0 re-tries
PAN card - fields, liveness, face match
Voice calls
The same loop with the screen switched off. Six fintech companies run live customer conversations on it today, inside their own infrastructure - and that is the phase that hardened everything above it.
See how it comparesThe same session, screen off
Ask for the lesson you want.
Pick a rung. It opens an email with the grade and subject already filled in - tell us who it is for and when suits, and we come back with a time.
Questions
The ones we are asked first. If yours is not here, any of the request buttons on this page opens an email - ask it in that.
What is BrainBack, exactly?
A live AI agent that teaches by talking and drawing. The learner joins a session in a browser and has a conversation with it. It draws on a board they both watch, hears the answers, and changes the lesson as it goes.
Is it live, or a recording?
Live, and both directions at once. The learner can cut in mid-sentence and the agent stops and listens - the same way a person would.
Who is it for?
Learners from class 1 to postgraduate, and the schools and edtech partners running them. The same agent also runs interviews, verification and customer calls for businesses.
How do I get a session?
Ask for one by email - the request buttons on this page open a message with the grade and subject already filled in. Tell us who it is for and when suits, and we come back with a time.
What do I need to use it?
A browser, a microphone and a reasonable connection. There is nothing to install.
What can it teach?
Eight subjects today, from class 1 phonics through to postgraduate cardiology. The lesson is generated for the learner rather than picked from a library, so the range is not a fixed catalogue.
What languages does it speak?
Almost any, in almost any accent, and it will switch mid-sentence when the learner does.
Can we run it on our own infrastructure?
Yes. Enterprise deployments keep every session inside the customer’s own environment - that is how the voice engine runs for its fintech customers today.