job-searcher

A twice-daily Telegram digest of fresh Product Design and UI/UX jobs — scraped from DOU and Djinni, delivered on time.

job-searcher watches two Ukrainian job boards, keeps only the Product Design / UI/UX vacancies posted in the last three days, and pushes a tidy, grouped digest to a Telegram chat every morning and evening. The same digest is archived as a dated Markdown file in reports/, so there is a permanent, browsable history. There are no servers to run and nothing to pip install: the scraping runs on GitHub Actions, and the punctual twice-a-day schedule is driven by a small Cloudflare Worker.

• Subscribe — open @JobsbroBot in Telegram and press Start. The current digest arrives within seconds, and the twice-daily one lands in your direct messages from then on.
• Digest — read the current digest in a browser, no subscription needed. Refreshed every half hour.

Sample output

Each run posts a message like this to Telegram — the same content is saved as a dated file such as reports/report-2026-09-26.md:

Design вакансії 🧑‍💻✨
🕒 26-09-2026 21:00 (Київ)
За останні 3 дні (24/26-09-2026)

🔵 Djinni (6/72) · 🗂️ Всі вакансії
⦿ 25-09-2026
• Product UI/UX Designer — Asocial · Remote
• UX/UI Product Designer — B2BSoft · Remote
⦿ 24-09-2026
• Product designer — bodo · UKR
• Product UX/UI Designer — In Sync · Remote
• Trainee/Junior UX/UI Designer — BasMar Software · UA
• UI/UX Designer — Vista Auction · Remote

🟢 DOU (3/261) · 🗂️ Всі вакансії
⦿ 25-09-2026
• Senior Product Designer — WhiteTech · віддалено
• UI/UX Designer — VCHASNO GROUP · Київ, віддалено
⦿ 24-09-2026
• UX Designer / Product Designer — Perla Helsa · Київ

Features

• Two boards, one digest. DOU and Djinni are scraped, normalized, and merged into a single message split into per-source sections, newest first.
• Only what's relevant. A title filter keeps Product Design and UI/UX roles and drops the rest (see Filter).
• Only what's fresh. A rolling three-day window by each board's own posting date, so you never re-read yesterday's list.
• Punctual delivery. A Cloudflare Worker cron fires within about a minute of the scheduled time; GitHub Actions' own scheduler was dropped because it ran best-effort and often fired hours late.
• Analytics over time. Each run records the day's scanned vacancy total per board — the count the board itself reports — and regenerates an Excel workbook with a line chart, committed to the repo (see Analytics).
• Almost no dependencies. The scraper and the report are pure Python standard library. The only dependency is openpyxl, used solely to write the Excel analytics workbook (the standard library cannot produce an .xlsx with a chart).
• Archived history. Every run lands a dated Markdown report through an auto-merged pull request, so the archive builds itself.

Analytics

Alongside the daily digest, every run records the scanned vacancy total each board carried that day — the raw count the board itself shows, before the relevance filter (so DOU's whole Design category and Djinni's Product Design + UI/UX tag listing) — and keeps a running time series:

• data/vacancy_counts.csv is the source of truth — one row per calendar day with the columns date, dou, djinni. A second run on the same day overwrites that day's numbers with the latest; a source that failed to scrape is left blank (a gap), not recorded as 0.
• Git activity — the repository's commit graph shows how the history builds up over time as each run lands its report and updated counts.
• Live chart on Google Sheets — a hosted view of the same data. A bound Apps Script pulls the CSV above from the repo on a daily trigger and redraws the chart, so the sheet tracks the history without a download.

Both files are committed by the same auto-merged pull request as the Markdown report, so the history builds itself. The workbook's document properties are set deterministically from the data, so a run whose counts did not change produces no spurious binary diff.

The bot's own audience is tracked separately, by the Worker:

• Subscriber dashboard — active and total subscribers over time, from one aggregate snapshot a day. It shows counts only, no chat ids, so the link is safe to share.

Run it locally

pip install -r requirements.txt
python3 daily_report.py

This writes reports/report-YYYY-MM-DD.md, updates data/vacancy_counts.csv and reports/vacancy-analytics.xlsx, and prints a short summary to stdout. Telegram delivery is skipped locally — it only sends when the TELEGRAM_BOT_TOKEN and TELEGRAM_CHAT_ID environment variables are set.

Running the tests

The parsing, report-rendering and analytics helpers are covered by an offline unit-test suite in tests/. The tests run against in-memory fixtures and never touch the network, so they are fast and reproducible:

pip install -r requirements-dev.txt
python3 -m pytest

The same suite runs in CI on every push and pull request (.github/workflows/tests.yml).

How it works

• scrapers.py fetches and normalizes vacancies from both boards using only the Python standard library. DOU is paginated through its xhr-load endpoint and its Ukrainian "day month" list date is parsed; Djinni is read from the JobPosting JSON-LD on each results page, including its datePosted. Only titles matching the Product Design / UI/UX filter are kept.
• daily_report.py runs the scrapers and, for each kept Djinni vacancy, reads the vacancy page's "Оновлено" (updated) date and uses it in place of the published date when present, so a re-bumped posting resurfaces. It keeps the vacancies whose date falls within the window (WINDOW_DAYS, default 3 = today and the two previous days), groups them by source, sorts newest first, stamps the generation time in Europe/Kyiv, writes the Markdown report, and sends the Telegram message(s). If a source fails, the report notes the error; if both fail it exits non-zero so no empty report is committed.
• analytics.py upserts the day's per-source scanned totals (the raw count each board returns before the relevance filter) into data/vacancy_counts.csv and regenerates reports/vacancy-analytics.xlsx (the Counts sheet plus a line chart) with openpyxl. See Analytics.
• trigger/ is a Cloudflare Worker that acts as the clock. Its cron calls the workflow's workflow_dispatch entry point on time — see trigger/README.md for a one-time deploy.

Automation

The report runs as a GitHub Actions workflow (.github/workflows/daily.yml). It is triggered twice a day — at 11:00 and 20:00 Europe/Kyiv, all year round — by the Cloudflare Worker in trigger/, which calls the workflow's workflow_dispatch entry point. The Worker's cron runs on UTC, so it picks the report ticks by their Kyiv time: 08:00 and 17:00 UTC in summer, 09:00 and 18:00 UTC in winter. GitHub Actions' own schedule: was removed: it ran on a best-effort basis and routinely fired tens of minutes to several hours late (or was dropped under load), so the punctual Worker cron replaced it.

Each run sends the digest to Telegram and, when the report changed, lands the Markdown file in reports/ through an auto-merged pull request on a short-lived branch rather than pushing to main directly. GitHub-hosted runners have the outbound network access the boards need; the scraping runs there because the Claude Code cloud sandbox blocks those connections by egress policy.

Telegram delivery is enabled by two repository secrets, TELEGRAM_BOT_TOKEN and TELEGRAM_CHAT_ID; without them the script just writes the report and skips the notification. TELEGRAM_CHAT_ID may list several comma-separated destinations (for example a personal DM alongside a channel).

Optionally the digest can run as a subscription bot: people open @JobsbroBot and press /start, and each receives the report in their own DM. Pressing /start also sends the current digest right away — the Worker serves it from its own storage in about a second, and keeps that copy no more than half an hour old with a refresh run, every half hour through the day, that scrapes and stores but sends nothing (rate-limited to one digest per chat every 10 minutes), so a new subscriber neither waits for the next scheduled run nor for a scrape. Subscribers are stored by the Cloudflare Worker in trigger/, and the report reads the active list (via the SUBSCRIBERS_URL secret) on top of any static TELEGRAM_CHAT_ID, retiring anyone who blocked the bot (via DEACTIVATE_URL). This is what makes the user count meaningful — see trigger/README.md for the setup and the /stats endpoint that reports how many people use the bot.

The digest itself is also readable as a web page, at /latest — the same vacancies the bot sends on /start, refreshed every half hour, so the link can be opened or passed on without subscribing.

A shareable dashboard charts active and total subscribers over time. Both pages show only what any subscriber would see — vacancies on one, aggregate counts with no chat ids on the other — so the links are safe to share.

Trigger a run by hand any time from the Actions tab (Run workflow) or with:

gh workflow run "Daily design vacancy report"

To put a one-off note above the digest's heading, for example to explain why a digest is being sent again, pass the notice input. Only that run's recipients see it; the digest the bot serves on /start leaves it out.

gh workflow run "Daily design vacancy report" -f notice="⚠️ Text of the note"

Filter

The relevance filter (in scrapers.py, RELEVANT) matches Product Design, UI/UX and Graphic Design titles in English and Ukrainian: product design, ui/ux, ux/ui, ux designer, ui designer, user experience, user interface, head of design (a leadership title that often names no specialty), vibe-code designer and vibe designer (a designer who builds with AI coding tools), graphic design (which also catches graphic designer), plus Ukrainian forms such as графічний дизайнер, продуктовий дизайнер, продакт-дизайнер, UX-дизайнер, UI-дизайнер and дизайнер інтерфейсів. The Ukrainian rules anchor on the specialty word, so a bare дизайнер is not matched. Adjust that regular expression to widen or narrow the report.

The digest itself is narrower than the scraper. RELEVANT is built from two patterns — PRODUCT_UI_UX and the graphic-design one — and daily_report._prepare keeps only the titles matching PRODUCT_UI_UX. That still matters for DOU, whose listing is the board's whole Design category and so carries graphic-design titles the digest drops. Djinni is scoped by tag instead, and DJINNI_URLS asks only for Product Design and UI UX, so graphic-design vacancies reach neither the digest nor the scanned total that feeds the analytics series. A title naming both families (Graphic/UX Designer) counts as UI/UX and is sent. Restoring graphic-design roles means putting the Graphic%20Design tag back in DJINNI_URLS and dropping the PRODUCT_UI_UX filter in _prepare.

A vacancy that asks for vibe coding stays where it is but gets a ⚡ vibe coding mark at the end of its line. daily_report.VIBE_CODING looks for the practice and the tools (Claude Code, Lovable, v0, Bolt, Figma Make, Replit, Windsurf) in the title and the description. Djinni descriptions arrive with the listing, while DOU needs one page load per vacancy in the window to read its description. The welcome digest sent on /start skips those page loads, so there a DOU vacancy is marked only when its title shows it.