← All work
GEO CONTENT OPS · AUTOMATION

Ceyuan · Content-Ops Console

Role
Solo · independent project
Status
Core workflow implemented
Stack
FastAPI · React · Postgres · Playwright · Doubao
A console for GEO content operations: business profiles, AI generation, human review and multi-platform publishing. Each step records its state, evidence and final URL for verification and manual follow-up.

Why I built it

As more people ask AI to find things, being recommended by AI becomes the new front door for customer acquisition.

I organized business profiles, content production and publication checks into one repeatable workflow. The engineering focus is on making generation, review and publishing inspectable, with explicit evidence and states at each step.

One console, the whole chain

Ceyuan is an operations console: build a client dossier, generate from it, QA by hand, publish to each platform. One screen shows where every client and every piece of content stands.

One screen, the whole chain: clients, content, every step from draft to published.
One screen, the whole chain: clients, content, every step from draft to published.

The hard part isn’t automation — it’s publishing itself

Publishing to these platforms is adversarial: some return a link instantly, some sit in human review, some block automation outright. The real risk isn’t failing to publish — it’s looking published when you’re not.

Ceyuan tracks each platform’s real state — published (with a real link), in review, needs a human — and re-checks after posting: no real link, not counted as done.

Publishing isn't one state: each platform tracked by its real result.
Publishing isn't one state: each platform tracked by its real result.
Re-checks the final link: no real URL, auto-demoted to manual.
Re-checks the final link: no real URL, auto-demoted to manual.

When a human steps in, it’s still on the record

Where automation can’t finish, a person does — and that step is logged too: who, when, which URL, note. No black-box step anywhere in the system.

Even the manual fallback keeps a full audit record: operator, time, final URL, note.
Even the manual fallback keeps a full audit record: operator, time, final URL, note.

Only the client’s own facts

Generation uses product information and images from the business profile, with prompts that instruct the model not to invent facts. Prompts alone cannot guarantee accuracy, so human review remains a required step. Drafts with missing evidence or unsupported claims need more source material, revision or rejection.

Generation uses the business profile as its source, followed by human review.
Generation uses the business profile as its source, followed by human review.

What I built

Implemented the profile → generate → review → publish workflow, including business profiles, content generation, human review, publication-state tracking and final-URL checks.