Slaydxona

Status: live. umbra.uz , with the bot at @SlaydxonaBot . A student types a topic and gets back a .pptx that’s ready for class: a proper university title slide, correct Uzbek (oʻ gʻ), short text and real sources.

Why it exists

Here’s the honest part: Slaydxona is an audience machine. I didn’t build it because the world needs one more slide generator. Every student in Uzbekistan has to make a mustaqil ish or taqdimot every few weeks. That’s a huge, repeating need, and it lives entirely in Telegram.

So it works like this:

The bot asks you to join the channel before the first deck. Referrals only count when the friend is new, joins the channel and actually generates a deck. The channel is the asset. When I have something new to show, Botxona first, those people already follow me.

Why Haiku makes it work

The free deck runs on Claude Haiku. It’s fast, it handles structured JSON well, and it’s cheap enough that a free deck costs me close to nothing. So “one free presentation” is a marketing budget I can actually pay. Sonnet is only used for paid Premium decks (speaker notes plus defense Q&A).

What keeps the cost flat:

• Images cost nothing. The model only returns an image_query; code fetches from Pexels, Pixabay and Wikimedia Commons. No AI image generation, which also keeps the decks from looking AI-made.
• Prompt caching on the system prompt and layout catalogue, and no thinking mode for slide text.
• One router, three providers. Haiku first, then GPT and Gemini as fallbacks behind the same JSON schema, so a 429 never turns into a failed deck.
• A daily spend cap set in the admin. When it’s reached, new decks are refused without using a credit, and I get alerts at 80% and 100%.
• Every request logs its tokens, cost and latency, so the admin shows cost per deck, not a guess.

Paid tiers are about 5 000 so’m (Standard) and 9 000 so’m (Premium). They’re paid by card transfer with a screenshot, the same flow as Botxona. Prices live in the admin, not in code.

Not AI slop

Students and teachers can spot an AI deck immediately: same layout on every slide, walls of text, a purple gradient. Before I designed a single template, I collected and analysed 201 real student decks (2 361 slides) from Uzbek universities and schools. I looked at how long they are, how many words go on a slide, what the title slide looks like and how they cite sources.

From that came 50+ master .pptx templates built on 25 real archetypes. The fonts are ones every student’s Office already has, so nothing shifts when they open the file. Sources follow the local academic style, and the title slide fills itself from your profile (university, group, “Bajardi”, “Qabul qildi”).

Users don’t pick a template. Out of 52 designs students chose badly (loud templates for law, chalkboard for medicine), so the AI classifies the topic and picks one. If you don’t like it, “Boshqa dizayn” re-renders the same content in another design, with no extra LLM call and no extra credit.

The admin is the real product

I spent as much time on the admin as on generation, because that’s where I actually learn. In one place I see:

• Errors: every failed generation, with provider, model, fallback and latency.
• Feedback: what users write back, tied to the deck it’s about.
• Analytics: users, decks, referrals, spend per day, payments queue.
• Broadcasts: segments with live counts, media, buttons, scheduling, pause/resume. It resumes after a restart, and nobody gets a message twice.

I don’t add features on a hunch. I read what broke and what people complain about, fix that, and ship it.

How it’s built (and kept)

Built in two days, 10–11 October 2026, with Claude Code, running the spec-first setup I keep improving:

• A CLAUDE.md with the decisions table. Anything in it doesn’t get re-argued without me.
• A STATUS.md handoff file plus /handoff and /resume commands, so any session can pick up exactly where the last one stopped.
• Subagents by cost: Haiku to explore and run QA, Sonnet to implement, Opus only for architecture review. The main session coordinates and doesn’t read big files itself.
• 733 tests, plus a template QA script that renders every template in Uzbek, Russian and English and checks glyphs and overflow.

Stack: aiogram 3, FastAPI, an arq worker, Postgres and React. It runs on the same box as Botxona with hard memory limits on the render worker, so it can’t starve its older sibling.

Results

• Most users are satisfied. The feedback in the admin is mostly positive, and the complaints are specific enough to fix.
• The channel grows by itself. Every free deck means one more channel member, and every referral means three.
• It’s paying for itself as marketing. Haiku keeps the cost per free deck tiny, and in exchange I get an audience of students and young developers, which is exactly who Botxona is for.

Users	—
Decks generated	—
Channel members	—
Avg. AI cost per free deck	—

Related: Ideas, Botxona, Telegram, AI agents, Spec-first delivery.