An end-to-end AI content pipeline, run live as a case study.
Wise Piece is a Traditional Chinese AI-content site producing 60+ articles on AI and automation pain points for Mandarin-speaking knowledge workers (PMs, designers, content creators, SME owners). Each article auto-generates: article body, cover image, 4 comic panels, 4 slide images, social media drafts — ships through GitHub → Vercel auto-deploy on an MWF cron.
I built it as a working experiment in what a solo non-engineer can ship with the current generation of AI tools. The site is both the deliverable and the case study — clients who want this kind of pipeline can see it running in production, including the parts that broke.
The end-to-end automation flow
Each article cycle on Wise Piece touches 8+ external services, orchestrated as one pipeline. This is the part that matters most about the system — it isn't a content site that also publishes to social; the multi-platform orchestration is the product.
- Trend detection — orchestrator scans signals and selects a topic (pluggable to RSS / SERP / social listening sources)
- Article writing — GPT-5.5 via OpenRouter (DeepSeek-V4 fallback on rate-limit / 5xx) generates body, summary, TLDR, direct-answer block, 2 FAQ items. Voice Guide loaded as system prompt for consistency.
- Image generation — KIE (gpt-image-2) produces 1 cover (16:9) + 4 comic panels (1:1). Wavespeed takes over automatically if KIE 429s, balance-dips, or returns garbled JSON. Brand-aligned placeholder ships if both providers exhaust.
- Slide rendering — 4 slide PNGs (1080×1080) rendered locally via Puppeteer from the article outline, uploaded to Cloudflare R2.
- Schema generation — Article + FAQPage + BreadcrumbList + Organization JSON-LD embedded in the page, with explicit hreflang for any translated versions.
- Web publish — article JSON written atomically (tmp + rename) to the monorepo, git commit + push to GitHub, triggers Vercel auto-deploy.
- Cross-platform social publishing via Meta Graph API — the same article auto-converts to:
- Instagram 8-image carousel (1 cover + 4 comic panels + 3 slide cards) via the Instagram Business Graph API
- Facebook feed post (article URL preview card + caption) via the Facebook Pages Graph API
- GPT-5.5 generates platform-native captions before publishing — IG hashtag-heavy, FB longer narrative, both linking back to the article. Auth via permanent Page Access Token.
- Newsletter dispatch — Resend transactional email to subscribers with article digest + cover preview. Welcome email includes IG + FB icon CTAs.
- RAG index refresh — embeddings.json rebuilt (~2 sec for 125 docs) so the in-site Paipai widget can retrieve the new article on the next cold start.
Total per-article runtime: 18-25 minutes from cron fire to live on site + IG + FB + newsletter + RAG. Each leg has fallbacks (LLM model fallback, image provider fallback, atomic checkpoint resume via RESUME=1). When the cron died in week 4, recovery was a single command that resumed from the last successful stage instead of starting over.
The stack
- Next.js 16 on Vercel · Cloudflare R2 for media · GitHub for source + auto-deploy trigger
- OpenRouter as the LLM gateway · primary GPT-5.5, fallback DeepSeek-V4-Flash (auto-switches on 429/5xx)
- KIE + Wavespeed for gpt-image-2 generation (primary + fallback, ~5–8 minutes per cover)
- OpenAI text-embedding-3-small for the RAG widget (in-memory cosine over 125 docs, 1536-D)
- Meta Graph API v25 (Page Access Token) for outbound IG carousel + FB feed auto-publish — every new article publishes to both platforms on the same cron
- n8n for inbound IG/FB webhook automation — comment-to-DM magnet flows, Data Table routing, magnet-aware reply rendering
- Schema.org JSON-LD for AI citation optimization: Article + FAQPage + BreadcrumbList + Organization + DefinedTermSet on the glossary
- Sharp + ffmpeg + Puppeteer for local media processing
Current numbers (6 weeks in)
- 63 articles published across 3 content pillars
- Every article auto-publishes to IG (8-image carousel) + FB feed + newsletter email on the same cron — no separate manual social workflow
- 62 AI-term entries in the glossary (with structured schema)
- 5-case portfolio · 1 real external client engagement
- ~$0.50–1.00 USD all-in cost per article (LLM + image gen + R2 storage)
- ~98% publish uptime after the Layer 1–3 fault tolerance work
- Google indexing ~16% in week 3 (new domain, AI content — in the standard observe-period range)
Available in English
This landing page, for now. English translations will be added selectively as we go — prioritizing articles that lean on primary research and concrete cases we've shipped ourselves, rather than auto-translating the full Chinese catalog. Quality over volume; more pieces to land here in the following weeks.
In Chinese only (for now)
- The full main site in Traditional Chinese (homepage, articles, pillar pages)
- The 62-term AI glossary with DefinedTermSet schema
- The portfolio of 5 cases, including a real travel-agency intel aggregation system
- Sharon's full bio (the tour-guide → AI translator story)
- Paipai (派派) — the in-site AI search widget, bottom right. Currently Chinese-only; ask it anything about the catalog
Who this is probably for
- Solo founders trying to build AI content automation without an engineering team
- Anyone working on GEO/AEO (Generative Engine Optimization + Answer Engine Optimization) and AI-citation tracking
- People curious about non-engineer PMs building production AI pipelines (what was hard vs what AI made easy)
- Anyone running content sites in non-English markets and looking at retrieval / schema patterns
Reach out
Main site (Chinese): wisepiece.com
Email: wisepiece111@wisepiece.com
Instagram: @wise.piece
Newsletter: subscribe (Chinese welcome email; the newsletter itself is currently Chinese-only)
Note: this page is the English entry point. The rest of wisepiece.com is in Traditional Chinese; we're translating articles selectively rather than mirroring the whole site automatically, which would degrade quality and trigger Google's scaled-content-abuse signals.
