GROUP A
Automation Systems
AutomationiGaming / Member Operations
Membership LINE bot: check-in, payouts, account binding, lucky draws — fully automated
Member campaigns ran entirely on manual labor: check-ins tracked in spreadsheets, payouts reconciled entry by entry by support staff, LINE identities never matching game accounts. More campaigns meant more misses and more mistakes.
- LINE-to-game-account binding: bind once and every later interaction resolves to the right member automatically
- Daily check-in rewards, lucky draws that draw themselves, automatic payouts — push notification the moment a reward lands, zero manual reconciliation end to end
- Campaign broadcasts and command queries (promos, progress); support staff handle exceptions only
- Deployed on a GCP cloud VM: systemd auto-start on boot + crash self-healing, external downtime alerts, database auto-backed up off-site every 6 hours
Check-in / payouts / binding / draws: all 4 automated
Database backed up off-site every 6 hours
24/7 downtime alerts + crash self-healing
Source: system deployment and runtime logs (cloud migration 2026-05)
AutomationiGaming / Community Operations
Community engagement bot: multi-persona group chat automation that doesn't read like a machine
A silent group is a dead group, and hiring part-timers to keep the chat alive is expensive and hard to manage: stories drift, shifts go unfilled, nobody remembers what was already said.
- Multiple fixed personas, each with a personality and long-term memory: yesterday's picks and win/loss records live in a ledger, so next-day recaps and cross-talk stay in character — no persona breaks
- All content from real data: fixtures, odds, and results wired to real sources, fabrication banned — call one game wrong and the group's trust is gone
- Human-like send cadence: multiple accounts interleaved, intervals randomly floating 10–15 minutes — like real people busy with their own lives, not bots flushing a queue
- Every message screenshot-archived and verified to have actually landed in the target group — delivery has an audit trail
4 fixed personas with long-term memory ledgers
Send intervals randomized 10–15 min, human-like
Every send screenshot-verified in the group
Source: sender runtime logs and per-message screenshot archive
AutomationSports Data / Content Media
A 400-plus-page sports content site with the whole update pipeline automated
400+ pages of match content change every day. Manually updating standings, scores, and fixtures daily was never going to hold.
- Whole site on a GCP VM; the fetch data → validate → compose → rebuild → deploy pipeline runs 4 times a day on its own
- Built-in numeric gate: standings, goal difference, and match counts cross-checked for self-consistency — fail the check, abort the run. Bad numbers never deploy
- Upstream source outages auto-retry with exponential backoff (15/30/60s); on failure the last good build stays live and an alert fires
402 pages generated automatically
Auto-updates 4x daily
0 bad-data deploys (gate aborts them)
Source: pipeline runtime and deploy logs
AutomationData Operations
Daily off-site backups for critical data, with three defenses against mirror wipe-back
The scariest backup failure isn't having no backup — it's a corrupted source getting mirrored outward and deleting the good copies with it.
- Scheduled daily incremental backups to off-site storage; drives identified by marker file, not drive letter — replug it anywhere and it's still found
- Collapse protection: an abnormal drop in source file count rejects the whole run; deletions are soft-deleted and kept 30 days; 30 snapshots retained
- Three-state Telegram alerts (halted / failed / recovered) — quiet on normal days
283 files SHA256-compared one by one, 0 missed
Incremental backup in ~1.3s
8/8 negative tests passed on the defenses
Source: backup system runtime and verification logs
AutomationContent Media / Sports Data
A content CMS you can publish from a phone — one tap, live in 10 seconds
The content team needed to file stories fast from a phone and go live with one tap — with zero contact with code or deploy commands.
- Minimal admin: brute-force-protected login, CSRF on every form, version snapshots, images auto-converted to WebP with EXIF stripped
- GCP VM + systemd auto-start on boot and crash self-healing; failures alert via Telegram automatically
- One-tap real publishing: tap → export → rebuild → deploy, end to end, fully automated
Publish end-to-end in ~10s
Killed process self-heals in ~6s
Security negative tests all passed
Source: live system tests and security test logs
AutomationConsumer / Event Marketing
Anti-abuse for a giveaway signup page, then reconstructing the real entrant list
A signup page's worst enemy is bots and the same person flooding entries — and afterward it's hard to tell which entries are real people.
- Five anti-abuse layers: duplicate blocking, IP rate limiting, unique keys, fingerprint hashing, human verification
- Multi-dimensional data cleaning: phone-number patterns, name collisions, number distance, IPv6 prefix clustering, fingerprint clustering — cross-checked
- Cross-referenced time cutoffs and test seeds to reduce the raw entry count to a credible list
142 total entries
~118 genuine after audit
Source: signup system database and cleaning audit logs
GROUP B
Ads & Tracking
Paid MediaE-commerce / Fresh Produce
A brand that closes sales off-site — pulling Purchase events back with CAPI
Customers arrive from ads, but purchases happen on a third-party order page off-site. The pixel never sees a sale, so the algorithm never learns who actually buys.
- Self-hosted Meta CAPI endpoint; browser Pixel and server send the same event_id in parallel, deduplicated automatically
- Off-site sales fed back as Purchase events; all match keys normalized + hashed server-side, bad values dropped automatically
- Offline events carry no IP/UA, keeping junk signals from dragging down match quality
event_id dedup verified live
Event match quality (EMQ) ~6–7
7 off-site sales fed back
Source: Meta Events Manager live checks and feed-back logs
Paid MediaHigh-competition Lead Gen
Don't take the ad platform's word for it — first-party counting that attributes every conversion
Ad platforms count non-ad conversions into their results, and the backend's signup records have no source field — ads, organic, and social all blur together.
- A Cloudflare Worker logs every impression and CTA click — no ad platform in the loop, immune to blockers
- Privacy-first attribution: only hashes stored, never raw IPs; new signups matched back against click records
- Daily report + 2-hour flash updates pushed to Telegram automatically
Proven on a 3-signup day: 2 ads / 1 non-ad
Cross-system hash consistency verified
Unit tests 11/11
Source: first-party counting database and match logs
GROUP C
Web Builds & SEO
Web / SEOE-commerce / Fresh Produce
A local grower's brand site: leave the payment flow alone, add the missing trust
The shop already had an ordering system and LINE. What it lacked was a brand site that gets seen and trusted — without building its own payment stack.
- Hand-built single-page brand site on Cloudflare Pages; orders stay on the existing system, payment integration skipped entirely
- All real photos converted to WebP; structured data, sitemap, and canonicals done right in one pass
- Google Business Profile linked to the site, product catalog listed, GSC verified and submitted
Image weight 38MB → 2.3MB (−94%)
CTR 9% at 28 days live
Average position 10.1
Source: Google Search Console (pulled 2026-07) and build logs
Web / SEOHigh-competition Lead Gen
A new site in a high-competition market, growing organic traffic from zero
Grow a brand-new site's organic search traffic from zero, inside a single hyper-competitive market.
- Static site on Cloudflare Pages; GA4 + GSC wired end to end, conversions counted via outbound-click events
- Long-tail content at scale + internal-link balancing, targeting "discovered but not indexed"; template dedup to avoid being flagged redundant
- Weekly performance comparisons pushed to Telegram automatically
Impressions 724 → 4,730 in 28 days (~6.5x)
Clicks 68 → 191 (~2.8x)
Index rate ~88% (153/173 pages)
Source: Google Search Console (pulled 2026-07, vs. May same year)
Web / SEOHigh-competition Lead Gen
Keyword cannibalization surgery: 35 terms fighting across 6 pages, reassigned to one term per page
One core keyword set was scattered across 6 pages competing with each other: the head term ranked 23.7 with no page dedicated to it, while the most-anchored page (33 internal links) got just 8 impressions at position 37 — its title never contained the term at all.
- Mapped the cannibalization with 4 weeks of real GSC query×page data, then reassigned page roles: head term, winning terms, and calculator terms each own one page — the money page's home turf untouched
- Found the anchor-text lever: related-article cards use the page title as anchor text, so changing 1 title fixed 26 internal anchors at once (0 → 26/26 containing the target term)
- Honest lastmod policy: added an updatedAt field so only genuinely edited pages get a new date — no site-wide date spam to game Google
- Orphan rebalancing: a ranking page cited 0 times went from 1 → 11 inbound links; 7 pages fighting over the brand term (3,088 impressions for 6 clicks) split up by intent instead
Internal anchors with target term 0 → 26/26
Orphan page inbound links 1 → 11
174-page audit: 0 duplicate titles / missing canonicals
Source: Google Search Console (pulled 2026-08) and verified build output
Web / SEOHigh-competition Review Content Site
Technical SEO on a review site covering 61 brands: from wasted crawl to fully indexable
Content volume was fine, but the technical layer was leaking: any bad URL returned 200 + the entire homepage (soft 404) and burned crawl budget for nothing, sitemap lastmod refreshed site-wide every day until Google ignored it, and internal links were monopolized by footer boilerplate pages.
- Fixed soft 404s at the root: bad URLs went from "200 + 87KB homepage" to a proper 404, concentrating crawl budget on real pages
- Rebuilt sitemap lastmod as a content-hash state machine: the date changes only when content actually changes, so Google trusts the crawl signal again
- Redistributed internal links: review-page inbound median 5 → 12, minimum 3 → 7; broke the boilerplate page's 66-link monopoly
- 27 titles exceeded the Traditional-Chinese SERP truncation length (~30 full-width chars) → all trimmed back, 0 pages over; plus a dead-link gate that scans every internal link pre-deploy and blocks on any target miss (negative-tested: it does block)
- GEO / AI-search groundwork: llms.txt generated automatically, rebuilt on every content update
Bad URLs 200 → 404 (crawl budget saved)
Inbound-link median 5 → 12
Overlong titles 27 → 0 pages
Source: build output and live verification logs (2026-08)
Web / SEOTech-fix Snapshots · Across Multiple Sites
SEO quick-fix log: pitfalls we've fixed with our own hands
Plenty of sites don't have a content problem — they're stuck on one invisible technical fault. Everything below is a real case, each with before/after measurements.
- New site with zero pages indexed → traced to canonical vs. sitemap URL formats at war (.html and extensionless 308-redirecting into each other); aligned all four spots + resubmitted the sitemap, cleared
- Ghost 404s in GSC (/cdn-cgi/email-protection) → caused by Cloudflare's email obfuscation rewrite; wrapped the source in email_off markers to fix at the root, 67-page site scan shows 0 remaining
- Duplicate pages cannibalizing → the same brand written up on two pages with contradictory scores; 301-merged + content consolidated, keyword infighting stopped
- Links in collapsed sections devalued → added a site-wide plain-text index on the homepage, covering all 16 "unrecognized URLs"
- Traffic-crash false alarm → built an 11-item exclusion SOP for cliff forensics: read weekly data not daily lines, and don't misdiagnose normal CTR swings
Source: per-site GSC and before/after fix records (2026)