Inspiration
In Uzbekistan a lot of paper, cardboard and metal could be sold to buyers, but people do not know what is accepted, how to prepare it, or what it is worth. That information lives in scattered OLX ads and Telegram groups, mostly in one language. We wanted one place inside the app everyone already uses: Telegram.
What it does
- Send a photo: Gemini identifies the material (13 categories, or "unclear"), says whether it is recycled in Uzbekistan, and gives preparation tips.
- Three languages: Uzbek, Russian and English. A new user gets their Telegram language and can switch any time.
- Telegram Mini App: a recycling map from OpenStreetMap, a prices table and a personal impact tab.
- Real prices: buyer prices from public OLX.uz ads in Tashkent, each with its ad ID and date. Prices older than 60 days are marked as possibly outdated. Materials without a clear per-kg price stay empty instead of guessed.
- Community: points for sorting (max 20 items a day count), reports of new points and prices that an admin approves with
/reviewbefore they go live. Telegram ids are stored only as HMAC-SHA256 hashes.
How we built it
With the Kiro IDE and spec-driven development: two specs (core-bot, 22 tasks; mini-app, 54 tasks), three steering docs, property-based tests with hypothesis, and Autopilot with permissions.yaml rules. We reviewed every spec before running it and wrote "Review corrections" into it (keyed hashing, daily point cap, stale prices, Mini App limits), so Kiro applied them in every task.
Stack: Python 3.14, python-telegram-bot 22, Google Gemini (google-genai), SQLite, OpenStreetMap Overpass, Leaflet, vanilla JS, GitHub Pages.
Full write-up of how Kiro was used: docs/KIRO_WRITEUP.md
Challenges we ran into
- All mocked tests passed, but the bot crashed on Python 3.14 because of an old python-telegram-bot. We added a startup smoke test (
python -m qayta --check). - A Cyrillic project path confused the agent's shell. A Latin-only folder junction plus steering rules fixed it.
- One Autopilot run paused because a sub-agent used the global Python. One steering line and a permissions file let the remaining 47 tasks run without prompts.
- Few buyers publish per-kg prices, so we only show prices with a clear figure, source and date.
Accomplishments that we're proud of
164 passing tests, a working bot and Mini App in three languages, and no invented data: every price points to a dated public ad.
What we learned
Reviewing the spec before running tasks is the cheapest place to fix design mistakes. Steering is also the right place for machine-specific rules, not just code style.
What's next for Qayta
Phone-verified prices from collection points in Tashkent and Ohangaron, a curated list of points (with their consent), and PET and metal prices.
