QuickPost — AI LinkedIn Content Assistant

Problem

Most professionals who want to post consistently on LinkedIn stall at the blank page — finding something worth writing about takes longer than writing it, so posting becomes sporadic instead of consistent.

Solution

Built a five-stage content pipeline — research, planning, writing, scheduling, and publishing — where each stage is a real Celery background job, not a single prompt pretending to be a workflow. Google Trends, curated RSS, and Reddit feed a Gemini-scored planning stage; drafts get an automatic revision pass when confidence is low; and publishing goes through LinkedIn's real OAuth and Posts API rather than simulating success.

Highlights

• Five-stage content pipeline (research, plan, write, schedule, publish) running as Celery background jobs with persisted state
• Real LinkedIn OAuth integration — encrypted token storage, real publish via the Posts API, engagement synced back
• Memory-aware planning: published posts are embedded with pgvector so new ideas are checked against real history
• Every AI call either returns a real result or a real error — no fallback path silently fakes data when a source or model can't deliver

Outcomes

• Full pipeline verified end-to-end against live Postgres, Redis, and external sources (RSS, Reddit, Gemini) in Docker
• Test suite passing against a live database rather than mocks
• Responsive UI verified at mobile, tablet, and desktop breakpoints
• Diagnosed and fixed subtle Docker-on-Windows networking and stale-dependency issues surfaced only during full end-to-end verification