Content AI Agent

This demo runs on two layers of architecture. The first is the agent itself that is built on n8n. The agent has access to tools/capabilities like image generation, video generation, image editing, multi-platform social copywriting and each capability is its own separate n8n sub-workflow. When a request is received, the agent reasons about which tool it actually needs to complete the request and calls the appropriate sub-workflows. For different sub-workflows, I explored different AI models to weigh real cost-versus-quality tradeoffs, rather than defaulting to a single provider. A nice byproduct of that search was ending up with a different provider for each capability (image, video, copy), which turned into a good hands-on practice for architecting around different model connectors and APIs.

The second layer was bringing that agent to the web, which came with a few practical frontend constraints.

• Async execution: Video generation takes 30 to 60 seconds, which easily hits browser and edge timeouts. Solved by setting up a submit-and-poll flow so the webhook acknowledges immediately while the server handles generation in the background.
• Session scoping: To prevent concurrent visitors from leaking context into each other's chats, I isolated conversation memory per user using a browser-generated ID.
• Format mismatches: Live traffic caught an edge case where browser-recorded audio defaulted to a video/webm container that Gemini's transcription endpoint rejected. Fixed by forcing the correct metadata in n8n before sending the payload.