Instagram AI Automation
Multi-model GenAI content pipeline
Five invented topics stand in for the real 100-topic database. Every output below is synthetic and deterministic (seeded), assembled in your browser.
- 1 · Topic selection
- 2 · Caption + hashtags
- 3 · Video generation
- 4 · Voice + assembly
- 5 · Publish
Press Run stage 1: Topic selection to walk the pipeline one stage per click, from topic to simulated publish.
Synthetic preview. No real data: no model is called and nothing is posted. Topics, captions, hashtags and the publish are invented in your browser. The stage order and tools (GPT-4o, Sora-2, Runway Gen-3, Luma, ElevenLabs, FFmpeg + Sharp, Instagram Graph API) are the audited pipeline's. Source repo is private.
Synthetic preview. No model calls and nothing is posted; stage order and tools (GPT-4o, Sora-2, Runway Gen-3, Luma, ElevenLabs, FFmpeg, Instagram Graph API) match the audited pipeline.
A personal automation that takes the author's own Instagram Business account from topic to published Reel with no human in the loop. The TypeScript/Node pipeline picks from a 100-topic database (10 categories) with a 30-day dedup history kept in a local JSON file, has GPT-4o write the caption and trending hashtags, generates 720x1280 vertical video through a provider cascade (OpenAI Sora-2 first, then Runway Gen-3, then Luma Dream Machine, chosen by available API key), layers an optional ElevenLabs voice-over plus background music, assembles and watermarks with FFmpeg and Sharp, and publishes via the Instagram Graph API's two-step create-container then media_publish flow, built to run on a daily cron. A companion app, a FastAPI backend on the OpenAI Agents SDK + ChatKit with a Vite + React 19 panel, exposes the workflow as agent tools. Scoped honestly: the Reel pipeline calls the OpenAI SDK directly (only the companion uses the Agents SDK), and the ~$0.81 per Reel is the repo's documented cost estimate, not a measured average.
- TypeScript
- Node.js
- OpenAI GPT-4o
- Sora-2
- Runway Gen-3
- Luma Dream Machine
- ElevenLabs
- FFmpeg
- Instagram Graph API
- OpenAI Agents SDK + ChatKit
Architecture · topic → published Instagram Reel
Topic selection
Picks from a 100-topic database (10 categories) with a 30-day dedup history.
Caption + hashtags
GPT-4o writes the caption and trending hashtags.
Video generation
OpenAI Sora-2 generates vertical video, cascading to Runway Gen-3 then Luma by available API key.
Voice + assembly
An optional ElevenLabs voice-over; FFmpeg + Sharp assemble and watermark the Reel.
Publish
Published via the Instagram Graph API (create-container → media_publish).
- AI stack
- GPT-4o · Sora-2 · ElevenLabs
- Topic database
- 100 (10 categories)
- Est. cost / Reel
- ~$0.81
- Publishing
- Instagram Graph API
What I'd improve
The weak points sit around the model calls, not in them: the Sora-2 step rides a beta endpoint behind a ts-ignore, Instagram needs a public video URL so the pipeline leans on free ImgBB hosting as its CDN, and the dedup history lives in a local JSON file that any fresh runner would lose. Next pass: a typed adapter over the three video providers, real object storage for media, durable state for post history, and the Graph API insights endpoint wired in so engagement is measured rather than assumed.