Favzon: a digital platform for selling perfume
A platform of four products: a mobile app with an AI assistant, a Telegram bot that takes orders on its own, the favzon.uz online storefront and a back-office panel for the shop team. Everything runs on one backend.



What's inside
About Favzon
Favzon is our studio's own product: a digital platform for selling perfume. It grew out of two problems. The first is choosing: you can't smell perfume through a screen, and shoppers don't search by notes but in plain words like “something light for summer”, “a gift” or “for an evening date”. The second is ordering: shoppers message the shop in Telegram, and a manager has to ask for the phone number, address and payment receipt by hand every time.
The platform is made of four products. Each does its own job, but they all share one catalog, one order database and one customer list. The idea, the interface design and the code for every part are ours.
Constraints
- The AI may only recommend fragrances the shop actually has: it must never invent names, prices or notes.
- Shoppers write in Uzbek, including dialect, Cyrillic script and typos.
- The cost of each AI reply has to stay under control: more conversations mean higher spend.
- Managers work in two places, a Telegram group and the back office, so an order's status has to match in both.
- The cart travels from the website to the bot inside a Telegram link, and that link can't be longer than 64 characters.
Favzon AI
A mobile app for iOS and Android. It puts an experienced in-store advisor in your phone: shoppers write in their own words, the AI clarifies when needed and shows 2–3 matching fragrances from the catalog, each with a reason.
Features
- AI chat: fragrance cards with a match percentage, rating, price and an add-to-cart button.
- Catalog: categories, search and filters; every fragrance shows its note pyramid, longevity and sillage.
- Factories: where each fragrance was made, details about the factory and the fragrances it makes.
- Sets: ready-made ones and sets shoppers build themselves.
- Cart, an address picked on a map and push notifications about order status.
- Three languages, light and dark themes.
How it's built
- Vector search finds, the AI chooses. The request is matched against the database by meaning, and the results are re-ranked. The model picks 2–3 of them and writes a one-sentence reason for each. The server checks every fragrance the model returns against the search results: anything that wasn't in them never becomes a card.
- A vague request gets a question first. For something like “I need a perfume for my mum”, the AI asks two questions: who it's for and which season. Answers are picked with buttons, and the question is asked only once per conversation.
- Replies stream in. Text appears word by word, and a status shows while the search runs.
- The model provider can be swapped. A cheaper model runs by default; if it doesn't respond, the request falls back once to another provider.
- Passwordless sign-in. A phone number and a one-time code, delivered via Telegram.
- A demo mode with no server. Every feature has a mock data source, so design and testing proceed in parallel with server work.
- Liquid Glass on Android too. We brought the glass-effect library into the project and fixed its broken edges on Android.
Favzon Bot
A Telegram bot that takes the order by chatting with the customer. The manager is left to check the payment, ship the order and answer questions.
For customers
- A customer coming from the website gets the bot with the cart already filled in: fragrance, size and price.
- The phone number is sent with one button, the region is picked from buttons, and the address comes as text or a location.
- After confirmation, the bot takes the payment receipt. Later updates (payment confirmed, on the way, tracking number, delivered) arrive in the same bot.
- Typing a fragrance name shows its sizes and prices, and the bot answers common questions itself.
- The menu has the catalog, cart, order history and a way to reach a manager. The bot speaks Uzbek, Russian and English.
For managers
- Every order arrives as a card in one Telegram group: fragrances, total, customer, address and a map pin.
- Under each card are Paid, Shipped, Delivered and Cancel buttons, and the card shows who pressed them.
- To answer a customer, a manager replies to their message and the bot delivers the answer.
- News: an admin sends a post to the bot, and the bot broadcasts it to every subscriber.
How it's built
- The cart lives inside the link. Each fragrance and size is written as a short code, so up to 10 fragrances fit into a 64-character link. The bot rebuilds the cart from the catalog and takes prices from the website.
- Address from a location. The bot turns coordinates into an address: district, street and building. The map service receives only the coordinates, never the name or phone number.
- Typo-tolerant search. A fragrance name is found even when it's written in Cyrillic or with a 1–2 letter typo.
- One order, one status. If two managers press a button at the same time, the status changes only once.
Favzon Website
favzon.uz is Favzon's online storefront. There's no sign-in and no payment on the site: shoppers pick a fragrance, and the order is placed in Telegram.
Features
- Catalog: search by name and notes; filters for gender, season, fragrance family, notes, size and price.
- Sets, and a page for every fragrance with its notes, sizes and prices.
- A “Find a fragrance” quiz: a few questions, then 3 matching fragrances from the catalog.
- Reviews, videos, FAQ, and delivery and payment terms.
- The Order button opens a manager's chat with a ready-made message in a regular browser, and Favzon Bot with the cart inside Telegram. Every order gets a code starting with #FZ.
- Three languages, light and dark themes.
How it's built
- Fast pages. Pages are built ahead of time and cached. When a price or text changes in the back office, the backend tells the website to refresh only that part.
- Hybrid recommendations. The quiz scores the site's fragrances by semantic similarity (embeddings) and simple rules. If the embedding service doesn't respond, the rules run on their own.
- Spam-free reviews. Anyone can leave a review without signing up: Cloudflare Turnstile keeps spam out, and a review goes live after an admin approves it.
- SEO and analytics. Separate pages for each language, a sitemap and social previews; GA4, Yandex Metrica, Meta Pixel and UTM tags.
Favzon Admin
The back office and CRM for the shop team. All the data the other three products show is managed here.
Sections
- Dashboard: today's orders and revenue, a revenue chart, orders by status and low-stock alerts.
- Catalog: fragrances, brands, factories, notes, categories and sets.
- Orders: from every channel — Telegram, Instagram, phone, website, app and offline sales.
- Customers: a customer card with phone, address, notes and purchase history.
- Website content: the home page, customer testimonials, videos and review moderation.
- Notifications, analytics, CSV import and user permissions.
How it's built
- Clear statuses. An order is Pending, Confirmed, Shipping, Delivered or Cancelled. A status set by mistake can be rolled back, and reopening a cancelled order checks stock again.
- Two-way sync with Favzon Bot. A button pressed in the group changes the status in the back office. When the status changes in the back office, the card in the group updates and the customer gets a message in their language. If the back office can't be written to, the order stays saved in the bot and the bot retries later.
- Secure sign-in. Staff sign in with email and password, tokens live in cookies that browser scripts can't read, and requests reach the backend through the panel's own server.
How the products connect
- One backend. Node.js and PostgreSQL: catalog, orders and customers in one database, images in Cloudflare R2.
- One order flow. The shopper chooses on Favzon Website and checks out in Favzon Bot, a manager confirms in the group, and the order is stored in Favzon Admin. At every step the customer gets a message in their language.
- One catalog. When a price changes in the back office, the website refreshes and the bot picks up the new price.
- One customer. A customer coming from the bot is matched to an existing customer card by phone, Telegram chat or username.
Where it stands
- Favzon Website is live at favzon.uz.
- Favzon Bot and Favzon Admin are built.
- Favzon AI is in development.
Website screenshots are from favzon.uz. Favzon Admin screenshots come from a test database with demo data: customers, orders and fragrance names are placeholders. Favzon Bot screens are a mockup drawn from the bot's real message texts. App screenshots are from its demo mode. Photos are from Unsplash.
Screens
Favzon AI

2–3 fragrances per request, each with a reason

For a vague request, the AI asks two questions first

Catalog: categories, search and filters

Fragrance page: note pyramid and character

List view: sets and fragrance notes

Factory page: details and its fragrances

Notifications: orders, discounts and recommendations

Menu: sections and chat history

Passwordless sign-in: phone number

The one-time code arrives via Telegram
Favzon Bot

Mockup: the cart from the website, price and free delivery

Mockup: phone, address and name, then confirmation

Mockup: order code, receipt and status updates

Mockup: the managers' group with card, buttons and reply
Favzon Website

Fragrance quiz: a few questions, then 3 matching fragrances

Ordering through Telegram in four steps

On mobile: picking the occasion

On mobile: picking favourite notes

On mobile: FAQ
Favzon Admin

Dashboard: today's orders, revenue and statuses

Orders: channel, status and payment at a glance

Order page and change history

Customer card: purchase history and average order
Need a product like this?
In a short call we'll discuss your idea and suggest the next step.

