Golang backend developer.

Problem

A logistics company was copying freight orders from load boards by hand into dozens of Telegram groups, order by order, operator by operator.

Architecture

A Go system that reads new orders from load-board bots in real time, parses route, cargo, weight and price, and hands each operator a ready card with price options — one tap publishes a formatted post to every one of their groups from their own account. Around it: bulk delete and re-post, a searchable order archive, personal filters, a night-time queue, statistics and scheduled reports, with multi-operator support behind owner approval.

The hard parts were not the features. Telegram stopped delivering login codes to the server IP, so login moved to QR; concurrent accounts needed live connections that recover on their own with per-account session state; rate limits needed a shared per-chat limiter with exponential backoff after 429; sessions are encrypted with AES-GCM, kept out of backups, and every action re-checks permissions.

load-board bots → parser → postgres → operator card → mtproto → telegram groups

Problem

A freight client needed new cargo listings on della.ua the moment they appeared. Watching the board by hand means missing them, and the same load surfaces more than once.

Architecture

A Go pipeline that monitors the board continuously and filters duplicates before anything downstream sees them. New listings go straight to a Telegram bot, and the parsed data is published asynchronously to Lardi-Trans through its official API. The whole application is driven from that same bot, including setting and changing the monitoring URL with pre-selected filters. Parsing started on Selenium and was later rewritten to plain HTTP requests — faster and far more stable. The Lardi-Trans integration became its own Go client library handling auth, request/response and error mapping, reusable across projects.

della.ua → http parser → dedup → telegram bot + lardi-trans api

Problem

Channel performance lives in YouTube's own dashboard, one channel at a time, with no history you control and no way to compare arbitrary periods.

Architecture

A Go backend that pulls data from the YouTube Data API on a configurable schedule and stores it in PostgreSQL, so history survives regardless of what the dashboard shows. On top of that: a video tracking table with view counts, views per hour and publish date, sortable columns and time-range filters from 24 hours to all-time, plus a channel module to add channels by username, hide them from tracking or drop them from the watchlist.

youtube data api → scheduled collector → postgres → tracking table → range & channel filters

Problem

Pool operators had to keep a dashboard open to notice a found block, a worker dropping offline, or hashrate moving — none of which happens on a schedule.

Architecture

A Go bot that polls the pool API and pushes to Telegram the moment something happens: a new block with the pool progress percentage at the time of discovery, a worker that stopped submitting shares, or hashrate crossing a configured threshold with follow-up alerts on every further step up. Block history is stored and readable inside the bot, with date, time and pool percentage for each one.

pool api → poller → storage → telegram bot → alerts + block history

Problem

Converting raster images to vector on vectorizer.ai in bulk, on a service with no official API, meant doing it by hand one file at a time.

Architecture

A Go bot that drives the browser end to end: upload, wait, download, repeat. Point it at a folder and get back the same number of vector files with no interaction in between. Batch size, delay ranges and pause intervals are configurable, with randomised gaps between uploads and scheduled long breaks between work cycles.

folder of images → browser bot → vectorizer.ai → vector files

Problem

Burning word-by-word animated subtitles into short-form video is the kind of work that eats an afternoon per batch when done in an editor.

Architecture

A self-contained Windows executable written in Go. Whisper transcribes offline, FFmpeg burns the result in, and the app handles word-level highlighting, 9:16 output at 1920×1080, and control over position, font size, line count and colours for text, active word and background. Drag in a folder, get the batch back. Nothing to install beyond the binary, and nothing leaves the machine after the model is downloaded.

drop folder → whisper (offline) → word timings → ffmpeg burn-in → mp4