Sadeek

Sellers had rough phone photos and no laptops. Sadeek uses AI to clean each image and create a complete listing for any product.

Putting a listing together still meant cleaning up the image, writing a description and deciding on a price. They already used Telegram, so I built the product there. A seller could send a photo and continue in the same conversation while Sadeek prepared the listing.

I built the whole product myself, including the bot, processing workflow, storefront and day to day operation. The storefront gave sellers somewhere to publish the result.

A useful listing needed more than a single model response. I connected background removal, relighting and AI try-on with descriptions, price estimates and moderation. The original workflow ran in n8n, with GPT and Gemini handling the content, pricing and moderation work.

I split processing into stages so I could check the result, retry a failed step and improve it without starting the whole request again. The bot also had to remember where each seller was in the conversation. Receiving a new photo, waiting for a result and managing an existing listing needed different responses.

1. 01

   Prepare the product image

   The input was often an ordinary phone photo. Image cleanup was part of preparing a listing, alongside the written content.

2. 02

   Review before publication

   I put moderation between generation and publication so the storefront would not publish every model response automatically.

3. 03

   Keep the flow inside Telegram

   I added request history, usage balances and subscription flows so sellers could follow their work and manage usage in the chat.

The next step was moving the proven n8n workflow into a complete backend that could handle the full product. It took a seller from a Telegram conversation through image cleanup, AI generated content and pricing, moderation, payment and a published storefront listing.

I built it with grammY, Bun and Express, backed by PostgreSQL, Redis and BullMQ. I separated conversation state, photo processing, payments and search into stages, so I could retry or improve one part without restarting the whole request.

Model routing, batching, caching and fallback providers helped keep paid API costs manageable and gave sellers a clear outcome when a service failed.

I built a Next.js storefront where buyers could browse the generated listings, search for a product and open a seller's store. Product pages brought the images, description, price and seller information together. I also adapted those pages for people browsing from a phone.

The storefront was the publishing part of the same workflow, rather than a separate product sellers had to fill in again. The archived screens below use demo listings.

The Telegram product was released. The later mobile app never shipped to the app stores.

I built the mobile version with React Native and Expo around an Arabic interface and right to left layout, with English and Kurdish translations as well. Sellers could select up to five photos and choose content generation, image editing or product identification. The app followed the queued work until the request succeeded or failed.

I also built product management, seller stores, a points wallet, price feedback, buyer and seller chat, and push notifications. Plan cards and splash screens could be configured from the backend without a new app release. The mobile screens shown here are from that unreleased version and use demo data.

Sadeek reached more than 10,000 users and handled more than 150,000 AI requests. It gave me experience building and operating an AI product on my own, where the cost of a request, the quality of the input and what happened when processing failed all mattered.

The project is neglected now. If I returned to it, I would decide earlier whether I was building a daily tool for individual sellers or a service other commerce platforms could integrate. I would choose the service for other platforms, set a cost budget for each request and introduce paid usage earlier. Free use helped me understand demand, but told me less about what people were willing to pay for.