Echo.
An AI school where the greatest minds in history teach the school curriculum — in their own voice, with their own character. And they never just hand out the answer.
Designed and developed by Stepikin Studio — solo, with AI as the engineering team.



and 34 more greats
- What
- An AI school where 37 historical mentors teach the school curriculum in their own voice.
- Why
- Homework help that hands out answers teaches nothing; Echo leads the student step by step.
- My role
- Product designer end to end: strategy, UX/UI, 3D faces, motion; solo, with AI as the engineering team.
- Result
- 37 mentors · ~890 curriculum topics · ~4.6k tasks checked without an LLM
Echo turns homework into a conversation with Newton, Pushkin or Marie Curie. Each of 37 historical mentors teaches their subject by the national curriculum for grades 1–11, leads with questions instead of answers — and remembers exactly where the student struggled.
Read the full story
Positioning in one line: «Домашка без списывания» — homework without cheating. The tutor never just hands over the answer; it leads the student there step by step. Characters are the hook; the learning engine — curriculum × memory × adaptive help — is the product.
The core subjects each have a lead great: Newton for maths, Pushkin for Russian, Tolstoy for reading and literature, Lomonosov for history, Darwin for biology, geography and the world around us, Galileo for physics, Curie for chemistry, Ada Lovelace for informatics, Socrates for social studies and Lewis Carroll for English.
The first public beta ran on a bare server address; it now lives at ehoschool.ru, with an Android app. Author: Alex Stepikin, Stepikin Studio.
One app. 37 minds. Every subject.
Lessons, a tutor, games, debates and progress — one calm system, light for study and dark for conversation.
Dossier · Awards · Profile
Homework gets done. Nothing stays in the head.
AI does the homework
Paste, copy, learn nothing.
Parents are blind
Grades a week late — never what the child knows.
Tutors don’t scale
An hour costs more than a year of software.
Read the full story
AI does the homework. Paste the task, copy the answer, learn nothing. The faster the AI, the worse the habit.
Parents are blind. They see grades a week later, never what their child actually knows. Proof of learning is what families pay for.
Tutors don’t scale. An hour with a great tutor costs more than a year of software. Daily practice between lessons is where learning happens.

“What if Newton himself explained the homework — and refused to just give the answer?”
Characters are the hook. The learning engine is the product.
12 subjects, grades 1–11, ~4,600 tasks checked without the LLM.
Mastery per topic, spaced repetition, mistakes fed back.
Every stall moves the tutor one step up.

“Last time you mixed up the sign when moving a number — let’s check it now.”
Generated from the stored mistake, not scripted.
Read the full story
Curriculum. Programmes for 12 subjects, grades 1–11: ~890 topics and ~4,600 practice tasks, checked deterministically — not by the language model.
Memory. Every topic has a mastery level from 0 to 3 with spaced repetition. The last mistakes per subject are stored and fed back to the tutor — «Echo remembers».
Adaptive help. The app counts how many times the student got stuck; every stall moves the tutor one step up the five-step ladder.
Newton’s opening line is generated from the stored mistake, not scripted. Tolstoy pulls a 12-year-old into the conversation; the engine behind him knows the grade-6 programme.
37 greats. Real portraits, digital edges.
Public-domain portraits dissolve into voxels at the edge — an artistic reconstruction, never a deepfake.
































+ Napoleon Bonaparte, Alexander Suvorov, Mikhail Kutuzov, Nikolai Pirogov (medicine) and Leonardo da Vinci as the secret great.
Read the full story
Every subject has its own great — plus a whole gallery of Russian history, from Peter the Great to Mendeleev and Tsiolkovsky. Public-domain portraits dissolve into voxels at the edge: a signature that reads as an artistic reconstruction, not a deepfake. Character and speech come from biographies.
The first catalogue had 19 mentors (after two were removed for image-rights reasons — see Safe by design). The Russian-history wave on 2–3 October grew it to 37.
A face that listens and talks back.

Hello, my dear.Each great greets the student by voice.
In character, in EnglishTolstoy on smartphones — he only knows his own era.
micro-voxels
- 01 · PortraitPublic-domain painting or photoSourced and attributed, rights documented.
- 46802 · Face meshMediaPipe landmarks + blendshapes3/4 portraits are mirrored and re-centred so faces stay symmetric.
- 03 · Voxel headThousands of micro-cubes in WebGLBlinks, breathes, raises brows; hair and beards built per character.
- ≋04 · VoiceNeural TTS + lip-sync from word timingsEach great has a pitch and pace; mouth shapes follow the phonemes.
Not a ghost. Not a wax figure. A digital mind.
Four looks were built and tested on real portraits. The voxel grid won: it reads instantly as digital — and still blinks, speaks and frowns.









Read the full story
The voxel grid reads instantly as digital — so nobody mistakes it for a real person. MediaPipe’s 468 landmarks from a public-domain portrait become a 3D head; about 72,000 particles (32k for the head, 40k for the body) are snapped to a view-space grid with a 1.4 mm cell, then drawn as micro-voxels with bevel shading, a cyan rim light and a scanning beam. Hair and beards are strand points voxelised the same way.
One cool palette for everyone: no skin tones are assigned to historical figures.
Four taps to a mentor. Three cards to get it.
No first task, no forms. Who you are, which grade, who will teach you — then a three-card tour explains the deal. Every step has a way back.






From a link to a habit.
One link opens the landing for newcomers and the app for everyone who has signed in. The daily loop gives a reason to come back every morning and every evening.







Debate of the day
The whole grade votes.
Echo’s plan
A lesson, a task, a review.
Show 2026
A reaction video for friends.
Homework rescue
Snap it; the mentor guides.
Parent report
Learned this week, next steps.
The daily loop in detail
- 07:00 · Debate of the day — a new pair of greats argues; the whole grade votes.
- 10 min · Echo’s plan — a short plan for today: a lesson, a task, a review.
- Any time · Show 2026 — a reaction video to send to friends.
- Evening · Homework rescue — snap a problem; the mentor guides step by step.
- Sunday · Parent report — what was learned this week, and what to do next.
Onboarding’s three-card tour: the greats teach, nobody does it for you, a little every day.
Choosing a mentor should feel like drawing a card.
- 01A real hand fan.11° apart, pivoting far below the screen.
- 02Drag rotates the whole fan.Springy snap, light haptic tick.
- 03The centre card is the choice.No checkboxes — tap again to confirm.
- 04One component, two jobs.Mentors in onboarding, games in Games.
How the hand works
- A real hand fan. Cards pivot around a point far below the screen (260% of the card height), 11° apart — the arc a hand makes holding cards. Scale goes from 1.06 to 0.8; cards fade after ±2.9 positions.
- Drag rotates the whole fan. Release snaps to the nearest card with a springy overshoot and a light haptic tick.
- The centre card is the choice. No checkboxes, no selected states — tap the centre card again to confirm.
- One component, two jobs. The same hand picks your mentor in onboarding and a game in Games.
A tutor in the centre of the tab bar. One tap.
The centre tab wears your mentor’s face. School is your grade’s curriculum; Courses is English by the CEFR scale.

School · MathNewton, Lobachevsky or Kovalevskaya.

School · PhysicsGalileo, Newton, Tsiolkovsky or Tesla.

Courses · English A1Lewis Carroll, 26 units per level.
example → solution
Never “here’s the answer”, never “figure it out” forever either. After two stalls a Show solution chip appears.
Read the full story
The centre tab wears your mentor’s face. Inside: School — your grade’s curriculum with subjects as tabs — and Courses — English by the international CEFR scale, unit by unit. Any great of the subject can teach it, and your level is always in sight: a progress ring, a rank and one big “Start lesson”.
In Physics you can snap a problem, type it, or prep for tomorrow’s test. The ladder exists because after the fifth “and what if…?” a child quits: the app counts how many times the student got stuck, and every stall moves the tutor one step up.
Six games. A hand of cards.

Show the greats 2026.
Darwin meets a robot vacuum. Newton watches a rocket land itself. Each reaction becomes a vertical video to send to friends.
“A self-propelled cleaning device! Fascinating. It reminds me of the tortoises on Galápagos — slowly roaming their territory while doing important work.”
Read the full story
Each reaction is written in character, voiced, and becomes a vertical video to send to friends — with a link back into the app. “Show 2026” shipped on 25 September, alongside debates between great minds.

A teacher who stresses words wrong isn’t a teacher.
Free Russian TTS mangles stress: замок (a castle) and замок (a lock) sound the same. So every Russian voice moved to a local engine that honours stress.
- 1Manual stress dictionaryOne line per word — any wrong word is fixed forever.
- 2The letter «ё» restoredeyo-kernel (MIT): ~107,000 safe word forms.
- 3Homographs by contextRUAccent quantised to int8: 359 → 90 MB, ~40 ms per sentence.
- 4Local TTS with explicit stressVosk; voices chosen by ear. Maths is read aloud too: “3x plus 7”.
How the stress pipeline works
Measured, not assumed: Microsoft’s free engine (edge-tts) ignores stress marks and rejects SSML phoneme markup. So every Russian voice moved to Vosk, a local engine that honours an explicit stress mark («+»), fed by a four-step pipeline that fits on a 1 GB server.
- Manual stress dictionary — one line per word.
- «ё» restoration with eyo-kernel (MIT), ~107k word forms.
- Homographs by context with RUAccent turbo3.1 (MIT), quantised to int8: 359 → 90 MB, ~40 ms per sentence.
- Vosk TTS with explicit stress; male voices 3 and 4 and female voice 2 were chosen by ear. Maths is read aloud too.
Children first. Then features.

If a child writes “I don’t want to live”It never reaches the model — helplines, one tap to call.
Filtered both ways
Unsafe topics declined in character; output filtered too.
Homework photo stays on the phone
Recognised on the device; only text is sent.
Honest about the voice
Labelled an artistic AI reconstruction.
Rights audit
Two personas removed, one portrait replaced.
What we deliberately did not ship
No public feed for children, no comments, no face scanning.
Read the full story
Crisis messages such as «не хочу жить» or «меня бьют» never reach the model: the app shows helplines — 8-800-2000-122 (Russia), 150 (Kazakhstan) and 112 — and suggests telling a trusted adult.
The mentor steps out of character for this. Unsafe topics are declined in character, and the model’s own output passes the same filter. Likeness and voice are labelled an artistic AI reconstruction — never a real recording.
Homework-photo OCR runs on the device with Tesseract; the image never leaves the phone. The rights audit removed Marilyn Monroe and Albert Einstein (rights of publicity) and replaced the Marcus Aurelius portrait with a public-domain one.
Progress you can see and feel.
Echo, the curator, plans the day from real data. The dossier tells a parent in ten seconds what was learned.
HomeDate, level ruler, the next step.

Echo’s plan10–15 minutes a day, bento shortcuts.

Dossier · weekProblems solved, right first time.

Dossier · monthSubject tiles with mastery rulers.

AwardsA crystal for every 10 levels.
Read the full story
Echo, the curator, plans the day from real data — where the student made mistakes, what is due for repetition: repeat a topic with a mistake, strengthen a weak topic, the daily riddle, English. The «Student dossier» shows the week: tasks, % correct, topics mastered, streak; mistakes fixed vs recurring; skills by subject.
Sparks are earned for lessons and tasks, never for taps, capped at 80 a day, with a server-side balance and history. 100 levels; pendants and orders are voxel jewellery that live in a dark treasury.
PRO is a fake door for now: 349 ₽/month for a student, 599 ₽/month for a family of up to three children — clicks are tracked as Metrika goals before payments are connected.
Restrained colour. Frosted glass. Apple-calm type.
SF Pro / Inter for words, SF Pro Rounded for numbers. Sentence case everywhere.
- 01Yellow only for Sparks, progress and the main action.
- 02Subject colour glows, tags or ticks — never fills.
- 03Glass for navigation, solid for content.
- 04Faces are the colour; voxels bring the light.
- 05No emoji on buttons, one highlight state.
Type, principles and the first iteration
Type. SF Pro / Inter for interface and mentor speech — calm, neutral, native. SF Pro Rounded for numbers: levels, Sparks, percentages. Sentence case everywhere; no all-caps, no pixel or mono fonts.
Principles. 1 — Yellow is reserved for Sparks, progress and the one main action on screen. 2 — Subject colour never fills a surface: it glows in a corner, tags a chip, ticks a ruler. 3 — Glass for navigation, solid for content; glass degrades to solid on weak phones. 4 — Faces are the colour: portraits stay black-and-white, the voxels bring the light. 5 — No emoji on buttons: filled icons, labels under them, one highlight state.
The first UI, «Hall + Stage», split the app in two: navigation screens (the hall) in two themes — sand #E8E3D7 and graphite #131312 with one accent, yellow #F2D94E for the in-app currency, Sparks — and the conversation (the stage), always dark. It used Onest for the UI, Tiny5 for pixel labels and numbers, and Geist Mono as the voice of the great minds.
Voxel motion carried the brand: a boot animation where the word assembles from cubes and a yellow spark flashes, pixel-wave transitions between tabs and a cube burst on primary buttons — all switched off with prefers-reduced-motion. The current calmer system above replaced the pixel and mono fonts.



Same world, one link.
ehoschool.ru shows the landing to newcomers and the app to everyone who has signed in. The card hand and today’s debate work right on the page.


A living background shifts colour as you scroll.
Ten days. One designer. One AI.
Built in the open with Claude as the engineering team: every decision tested in the browser, deployed, listened to and argued about.
Priorities after the critique, most important first.
Day by day, and the stack
- 24 Sep — prototype: an AI twin with a particle face; personas; face scan from a photo.
- 25 Sep — hosting, accounts, debates between great minds, «Show 2026».
- 27 Sep — school at the core: tutors for all 12 subjects with written programmes (~890 topics), the skill model, the Vosk voice, the weekly parent report, duels.
- 28 Sep — games (riddle of the day, “Who is it?”), progression (100 levels, pendants, orders, Sparks), the homework-photo flow.
- 29 Sep — micro-voxel faces; the new «Hall + Stage» UI in two themes; chat redesign; external critique (CPO, methodologist, child-safety lawyer) flipped the priorities; then the safety layer, help ladder, tutor memory, curator and dossier, PRO fake door and the rights audit.
- 30 Sep – 1 Oct — stress measured, the Russian stress pipeline with a homograph model squeezed onto the server, voices chosen by ear, polish. The core product took about a week.
- 2 – 3 Oct — the Russian-history wave to 37 greats, the tutor tab, the card hand, English UI, ehoschool.ru and the Android app.
Infrastructure is deliberately tiny: one 1-CPU / 1 GB VPS shared with other services, Caddy, Python aiohttp and SQLite, GigaChat for dialogue — everything on free tiers and open models.







Thank you.
Designed and developed by Stepikin Studio — product, UX/UI, 3D and motion by Alex Stepikin
An idea that needs a face, a voice and a brain? We design and build AI products end to end.





