maymun

Status: designing. The decisions are locked, the engine is specced, there’s a prototype page — and deliberately no product code yet, because one thing has to be proven first. More on that below; it is the interesting part.

The pitch, in one sentence

Savolingizni jonli darsga aylantiradi. — it turns your question into a living lesson.

You ask something, in Uzbek, Russian or English. Seconds later you’re watching a lesson: a real voice explaining it, with shapes, diagrams, code and maths appearing in time with the narration, and a pointer resting on the exact thing being talked about at the exact moment it’s said. You can pause, run the code, answer a checkpoint, or ask about any single moment and follow it deeper.

Nothing is a video file. A lesson is data — slides, narration, cue timings — played by a player I write.

What it is not

Writing this list down was the most useful hour of the whole design:

• Not a course platform. No syllabus, no enrolment, no fixed progression. Open topics and follow-ups; you file what you learn into your own tree.
• Not a chatbot. There is no chat window as the main surface.
• Not a video site. See above — nothing is rendered to a file.
• Not “ChatGPT with slides.” That phrase is the failure mode. Every roadmap item gets judged on whether it strengthens the synced engine or dilutes it into that.

Why Uzbek

Because nobody builds for it. Every tool my peers use explains the world in English or Russian, and the ones learning in Uzbek get the worst version of everything. An Uzbek-first explainer isn’t a translation feature bolted onto a product built for someone else — it’s the product.

That’s also where the moat is, if there is one: Uzbek narration plus animation plus the pointer, all in sync, plus a memory of what you’ve already learned.

The one thing that decides it

Not the AI. The voice.

The entire idea rests on narration that sounds like a person and timings precise enough for a pointer to follow the words. And that is exactly where the ground is soft: no major vendor does Uzbek well, and OpenAI’s TTS returns no word-level timestamps — which is the thing the sync depends on.

So the plan starts with a hard gate, before any product code:

1. Benchmark OpenAI, ElevenLabs, Google, Azure and self-hosted XTTS-v2 on real Uzbek text.
2. Derive word timings with forced alignment (WhisperX / MFA) rather than hoping a vendor hands them over.
3. Blind-test with ten native Uzbek speakers. If the voice doesn’t pass, there is no product, and I want to learn that in week one rather than in month five.

The second spike is nearly as sharp: the player assumes a browser, but the product is mobile-first React Native. One narrated slide with pointer sync, built both ways, measured on a ~$150 Android — see React Native. If it can’t feel good there, the whole mobile-first decision is wrong.

Constraints I chose on purpose

• No AI-generated images in v1. The “alive” feeling comes from synced SVG shapes, diagrams and the cursor — deterministic, instant, free, never wrong. Generated pictures would add cost, latency and a new way to be confidently incorrect.
• A flat price per lesson, tiered by depth. The user always knows what a thing costs before spending; the cost variance is my problem, not theirs.
• Bootstrap budget. No Kubernetes, no GPU fleet unless the voice forces it, managed services only. Every architectural choice gets judged on operational burden first and elegance second.
• Closed beta ≈ 100 invited users, engine and app only. The economy, payouts, payments, the public feed and B2B all wait.

What actually worries me

The brief describes about four products: an engine, a mobile app, a token economy and an admin studio. Solo. Scope creep is the default failure mode here, not a risk I might encounter — which is why the beta scope is written down as a fence rather than a goal.

The other one is honest bandwidth. shoshshi is live and wants attention, dona.uz  is my job, and this is the biggest of the three. Something has to give, and pretending otherwise is how all of them end up half-built.

Stub. Phase 0 results go here — what the ten listeners actually said, and whether the pointer stayed on the word.