🤖 Umeds AI Automation

Read keyword research database, pull Google Search Console data (clicks, impressions, positions), scan 20 RSS health feeds, check Google Autocomplete suggestions via Brave Search API.

Apply priority matrix: highest search volume + lowest competition. Check against published-articles.json to avoid duplicates. Verify topic doesn't exist on umeds.id via site:umeds.id search.

Run gen.py with topic-based prompt → Alibaba Wan 2.7 Image Pro generates medical illustration → saved as .webp file. If image fails, entire pipeline stops.

AI writes ~1,200 words in Bahasa Indonesia. Structure: H1 → H2 → H3 with FAQ section from Google Autocomplete. Internal links verified from sitemap. Medical authority tone for dental students.

POST to Umeds API with title, HTML content, description, and featured image (multipart/form-data). Article saved as draft for Minhaj to review before publishing.

Send summary: chosen keyword, article title, article ID, and 2-4 bullet points explaining why this keyword was selected (volume + competition data).

Run news-search.mjs → scrapes RSS feeds (Detik Health, Kompas, CNN Health, etc.) + Google News medical keywords. Aggregates and ranks topics by relevance to dental/medical students.

AI filters for dental student relevance: exam news, career updates, clinical cases, ethics stories. Picks the single best topic that hasn't been covered yet.

gen.py creates a relevant news-style illustration using Alibaba Wan 2.7. Pipeline stops if image generation fails.

AI writes ~800 words in news format: lead paragraph first (who/what/when/where), shorter paragraphs, factual tone. Targeted at calon dokter perspective.

POST to Umeds API under the News category. Article goes live for review.

Article ID, title, slug, and source summary sent to Minhaj.

Connect to Google Search Console (sc-domain:umeds.id) for clicks, impressions, CTR, positions. Pull GA4 data (property 327027843) for users, sessions, pageviews, bounce rate, traffic sources.

Run RSS trends script for health/dental topics. Refresh sitemap URLs from umeds.id/sitemap.xml for internal linking verification.

Pull latest code from GitLab (ummacademy_client). Scan for: missing alt attributes, hardcoded localhost, missing lazy loading, broken schema markup, meta tag issues, render-blocking resources.

Create Git branch seo/auto-fix-YYYY-MM-DD → apply fixes → commit → push → create Merge Request on GitLab. If no issues, skip this step.

Markdown report with all metrics, low-CTR opportunities, and findings → converted to PDF via Puppeteer.

Send PDF file to team WhatsApp group with brief caption. If MR was created, follow up with MR link.

Pull this week's GSC data (clicks, impressions, average position) and GA4 organic sessions. Compare against previous week.

Compare week-over-week changes against configured thresholds from penalty-monitor.config.json.

AI evaluates red flags:

✅ All Clear → Telegram: "No penalty signals detected" with metric summary
🚨 Alert → Telegram: Red flag details + auto-recommend pausing article generation pipeline

Read published-articles.json for articles from this week. Cross-check with sitemap for any missed. Check published-emails.json to never re-feature an article.

AI assembles newsletter from dokter POV (casual tone). Structure: greeting → featured article (dark card) → 2-3 more articles (green border cards) → clinical tip box → closing. CTA buttons link to articles.

Casual, engaging subject from dokter perspective with emoji. Example: "Pasien tanya scaling sakit nggak, Dok? 🦷"

POST to Umeds Email Blast API with status='publish' → immediately queues and sends to all users.

Subject line, articles featured, email blast ID, recipient count sent to Minhaj.

Python script connects to Umeds MySQL database. Runs team-specific queries for revenue, product performance, user churn, and conversion metrics.

Build targeted insights for 6 teams: CEO, Tech, Marketing, Product Learning, Finance, Design Creative. Each section has relevant KPIs and actionable recommendations.

Convert team-specific report into a polished PDF document.

PDF file sent to team WhatsApp group with brief caption and date range.

Pre-scheduled one-shot cron job triggers at exact time. Two emails: Heavy Reminder (Jun 5) and Last Call (Jun 7).

Read the pre-designed email template from ./emails/66-flash-deals-*.html with urgency-driven copy and flash deal pricing.

submit_email.mjs sends to Umeds API with status='publish' → emails queued and sent immediately to all users.

Confirmation sent to team WhatsApp group. Cron job auto-deletes itself after successful run.

Read all daily memory files (memory/YYYY-MM-DD.md) from the past 30 days. Each contains conversation logs, decisions, and context snippets.

Each memory entry is scored based on: recall frequency (≥3 times), semantic similarity score (≥0.8), unique query diversity (≥3 queries), and recency decay (14-day half-life).

Top 10 qualifying entries are appended to MEMORY.md — Sera's persistent long-term memory. These survive across sessions and inform future decisions.