👗 GAN-MNIST-Fashion: DCGAN for generating clothing images from Fashion‑MNIST.

Compared classic vs improved DCGAN: the classic trains faster, the improved captures finer details.
Implemented a custom GradientTape loop, added augmentations, and visualized loss dynamics.

📊 Results include epoch comparisons and generated samples.

🐱 Repo with code:
github.com/SC1DR-OFFICIAL/GAN-MNIST-Fashion

💫 Telegram Stars Invoice Bot: issue invoices in Telegram Stars with a single command.

I often get orders from abroad and noticed that many clients find crypto payments inconvenient.
I looked for an alternative and found a perfect one — Telegram Stars, which are easy to buy right inside Telegram anywhere in the world.

But Stars (now) can’t be directly transferred to another person.
So I wrote a bot that lets you instantly issue an invoice to a client in Telegram Stars.

Thought this tool could be useful for others too, so I released the source code on GitHub under an open license.

🐱 GitHub repo:
https://github.com/SC1DR-OFFICIAL/TG-Star-Invoice-Bot

🎮 I trained a neural network to play Pac-Man using 4 Reinforcement Learning methods

🐱 More details in my GitHub repo:
https://github.com/SC1DR-OFFICIAL/Simple-RL-for-Atari-MsPacman

🤖 I trained a neural network with PPO to control a robotic manipulator in the PandaReach-v3 task

🌪 The project also explores performance under uncertainty — when observations and actions are affected by noise

🐱 Results, graphs, and animations on GitHub:
https://github.com/SC1DR-OFFICIAL/ppo-panda-uncertainty

📱 I participated in the development of a multi-platform application for communicating with AI using Flutter.

• Secure two-step authentication (API key + PIN)
• Support for OpenRouter and VseGPT keys
• Local storage of history and analytics
• Token and spending tracking

🐱 More details in my GitHub repo:
https://github.com/SC1DR-OFFICIAL/AIChatFlutter/tree/main

👋 everyone!
I’d like to share a few highlights from the past week.

First of all, I finally completed the “Neural Network Developer” course. Over the past year, I invested more than 200 hours into it, and successfully passed the final assessment — a big and meaningful milestone for me.

Secondly, I took part in a hackathon for the first time, where we had to build a RAG system. Although we didn’t make it to the finals, my team still ranked 32nd out of 180 teams — a solid result for a first attempt. Most importantly, I met some great, intelligent people and significantly improved my understanding of RAG architectures.

Now that I have more free time, I’m planning to dive deeper into work and new projects! 🚀

🏚 Auto-publishing ads on OLX

For one of my projects: a bot for selling real estate in Tashkent the customer asked to add auto-publishing of ads on OLX.
The task turned out to be more difficult than it seemed. OLX does not have an API for working with an account, so I had to use browser control. But OLX has protection against bots: it recognizes work on the server (no screen) and blocks.
I figured out how to bypass the protection and even created an AI to pass the captcha if it appears.

Now ads from the bot are automatically published on OLX without manual intervention.

👍 Tired of Bitrix and heavy CRMs?

I build custom CRMs inside Telegram — with any functions you need, including AI (summaries, auto-tagging, lead scoring, etc.). For a real estate client I made a bot: client leads, exclusive listings with photos and listing IDs from the ads board, search by CRM or ad board ID, comments, and notifications to the team. No separate site, no Bitrix.

Want a CRM that fits your process and can use AI? I can do the same for you. 🌠

👋 Hello everyone!

For the past six months, I've been working primarily on large projects, including large commercial services, as well as research at the intersection of botany and artificial intelligence. I'm now approaching the final stages of each of them, and I've decided that writing about the intermediate stages wouldn't be as interesting, so I'll be covering them in detail soon.

🏚 I created a real estate ecosystem in Tashkent for a real estate agency, from client searches to internal team tools.

🚀 For clients, there's a Telegram bot with filters, subscriptions, and current listings in the city center. For the team, there's a CRM, property tracking, and daily processes. OLX monitoring is also a separate feature: new and updated listings arrive quickly, according to the required categories and tags. Posting is seamless—listings can be posted in channels and on OLX without manual copy-pasting.

🤖 The main feature is the AI secretary. It monitors incoming listings in messages, helps sort and organize them, assembles a proper listing from the chaos of messages and photos, asks clarifying questions, detects duplicates by photo, and then manages the cycle: publishing, editing, and deleting—simultaneously in the database, channels, and on OLX. It's not just a "smart chat," but an assistant that truly reduces the burden on the agency's day-to-day work.

Bot link

🙏 I built the technical side of cmdSAVE — a Telegram media bot that turns links into ready files: video, music, stocks, 3D, and more. My friend and motion designer Dmitry Tolstykh invited me into the project; he covers the product story in @nonasosal.

🟪 For users, it’s simple: send a link, pick quality or let the bot choose, watch live progress, get the file. It covers social video, music (tracks, albums, playlists), stock platforms, Behance, and Sketchfab 3D models. Plus adds playlists, priority processing, large-file delivery, and stocks without per-file packs — paid via T-Kassa and Telegram Stars. There’s also RU/EN UI, a status page, and a support bot.

🟪 Built as a real production service, not a thin wrapper around a CLI tool: resilient downloads when sites change or break, browser-assisted flows where needed, smart format selection, delivery past Telegram size limits, a priority task queue, billing, and health checks. Docker, Postgres, Redis — the usual adult stack for something that has to stay up.

The bot is currently in open testing, and any potential issues will be debugged. There are also plans to expand the service by adding books, movies, and TV series, and to create a website.

Telegram Bot · Status · Support