How to build Productized AI Systems (Content System)

Key Concepts

• Productized AI Systems: AI systems built once and resold to multiple clients without customization, automating end-to-end workflows.
• Resellable AI System: Pre-built system usable by multiple customers without customization, with personalization built into the system.
• End-to-End Workflow Automation: Automating as much of the workflow as possible to increase the system's value and selling price.
• User Interface (UI): A user-friendly interface for non-technical users to manage the workflow easily.
• Air Table: Used as a database to store and organize data.
• n8n: Used to automate the workflows.
• Air Table Interfaces: Used to create a user-friendly interface for the customer.
• Human-in-the-Loop: Allowing human intervention and adjustments in the AI-driven processes.
• LinkedIn Content System: An example system built to generate LinkedIn posts and images, adapt to tone of voice, use writing frameworks, and manage a content calendar.
• Campaign Setup: Personalizing the system for each company, generating business context, and scraping reference posts.
• Content Generation Request: Requesting new LinkedIn posts by repurposing blog articles or YouTube videos.
• Posting Frameworks: Different writing styles for LinkedIn posts, such as story-based, insight-based, and engagement-based.
• Analytics: Tracking post performance to create self-improving systems.

LinkedIn Content System Demo

The video demonstrates a LinkedIn content system built to manage content generation for multiple businesses. The system includes four sections:

1. Campaign Setup: Manages different companies, personalizing content based on company information.
   • Includes company name, URL, LinkedIn URL, and reference LinkedIn URLs.
   • Uses an AI agent (Perplexity) to generate a business overview, ICP (Ideal Customer Profile), and value proposition for each company.
   • Scrapes reference posts to mimic tone of voice and style.
   • Example: Scraping data for "Beni Agency" and "HubSpot."
2. Analytics: Checks analytics for past posts and references specific posts to mimic tone of voice.
3. Request New LinkedIn Posts: Generates new posts based on YouTube links or blog links.
   • Allows users to select a CTA (Call to Action) and add custom instructions for post and image generation.
   • Uses an L&M chain to generate three types of LinkedIn posts: story-based, insight-based, and engagement-based.
   • Uses the OpenAI image generator API to generate images.
4. Content Calendar: Schedules posts for specific dates.
   • Posts can be automatically posted on LinkedIn (though manual posting is preferred).

Building the Productized AI System from Scratch

Requirements for a Productized AI System

1. Resellable AI System: Pre-built and usable by multiple customers without customization.
2. End-to-End Workflow Automation: Automate as much of the workflow as possible.
3. User Interface: A user-friendly interface for non-technical users.

Initial Idea of the LinkedIn Content System

1. Strategy (Posting Strategy Setup): Personalization for each company, adapting to tone of voice, business, and ICP.
2. Input: Repurpose blog articles and YouTube videos into LinkedIn posts.
3. Post Generation and Scheduling: Write posts, generate images, include tone of voice, business context, and posting frameworks.
4. Analytics: Track post performance to create self-improving systems.

Building the System

• Tools: Air Table (database), n8n (automations), Air Table Interfaces (UI).
• User Flow:
  1. Main dashboard overview.
  2. Company/Campaign setup (add new companies, personalization).
  3. Scrape reference posts (mimic tone of voice).
  4. Content generation request.
  5. Human review and scheduling.
  6. Content calendar.

Back End Automations

1. Company Setup:
   • Generate a business overview and ICP for the company.
   • Scrape reference posts and their analytics.
2. Content Generation:
   • Repurpose YouTube URLs to LinkedIn posts.
   • Repurpose blog URLs to LinkedIn posts.
3. Scheduling:
   • Automatically post on LinkedIn at the scheduled time.

Setting Up Air Table

• Create a new base called "LinkedIn Agency in a Box."
• Create tables: "Company Setup" and "Select Reference LinkedIn Posts."
• Use Claude with the Air Table MCP to add fields to the tables.

Company Setup Table Fields

• Company Name (text)
• Company URL (URL)
• Company LinkedIn URL (URL)
• User LinkedIn URL (URL)
• Reference LinkedIn URL (URL)
• Reference LinkedIn URL 2 (URL)
• Generate Information (checkbox)
• Business Overview (long text)
• ICP (long text)
• Value Proposition (long text)

Select Reference LinkedIn Posts Table Fields

• Creator Profile URL (URL)
• Post URL (URL)
• Post Content (text)
• Likes (number)
• Comments (number)
• Select (checkbox)
• Image URL (URL)
• Company (link to company setup)

Setting Up Air Table Interfaces

• Create a new interface called "LinkedIn Agency in a Box."
• Use the "Record Review" layout for the campaign setup page.
• Use the "List" layout for selecting reference posts.
• Add a button to add new companies using a form.

Automation 1: Generate Business Overview and Scrape Reference Posts

1. Trigger: When "Generate Information" checkbox is checked in the "Company Setup" table.
2. Action: Run a script to trigger a web hook in n8n.
3. n8n Workflow:
   • Web hook trigger.
   • Air Table node to get the record information.
   • AI Agent (Perplexity) to generate business overview, ICP, and value proposition.
     • Uses company name, URL, and LinkedIn URL as input.
     • Uses a prompt to instruct the agent to generate the required information.
     • Uses an output parser to extract the three data points.
   • Relevance AI tool to scrape reference posts from LinkedIn URLs.
     • Transforms the LinkedIn URLs into an array.
     • Uses an HTTP request to call the Relevance AI API.
   • Split Out node to split the LinkedIn posts into individual items.
   • Air Table node to create a new record in the "Select Reference LinkedIn Posts" table for each post.
     • Links the post to the original company.
   • Air Table node to update the "Company Setup" table with the business overview, ICP, and value proposition.

Automation 2: Request New LinkedIn Post

1. Trigger: When a new record is created in the "Request LinkedIn Post" table.
2. Action: Run a script to trigger a web hook in n8n.
3. n8n Workflow:
   • Web hook trigger.
   • Air Table node to get the record information.
   • Switch node to route the workflow based on the type of repurposing (YouTube or Blog).
   • YouTube Repurposing:
     • HTTP Request to Rapid API to get the YouTube transcription.
     • Split Out node to split the transcription into individual sentences.
     • Aggregate node to combine the sentences into a single text.
     • Set Fields node to convert the text into a string.
   • Blog Repurposing:
     • HTTP Get Request to scrape the blog content.
   • No Operation node to combine the two workflows.
   • Air Table node to get the company information (business overview, ICP, value proposition) from the "Company Setup" table.
   • Air Table node to search for the selected reference posts in the "Select Reference LinkedIn Posts" table.
     • Filters the posts based on the company and the "Select" checkbox.
   • Aggregate node to combine the post content into a single list.
   • Set Fields node to convert the list into a string.
   • L&M Chain to generate three different styles of LinkedIn posts (story-based, insight-focused, engagement-focused).
     • Uses the content (YouTube transcription or blog content), company information, and reference posts as input.
     • Uses a prompt to instruct the agent to generate the posts.
     • Uses an output parser to extract the post title, content, type, and suggested hashtags.
   • Split Out node to split the generated posts into individual items.
   • L&M Chain to generate a prompt for the image generator.
     • Uses the post title and content as input.
     • Uses a prompt to instruct the agent to generate the image prompt.
   • OpenAI Image Generator to generate an image for each post.
     • Uses the generated image prompt as input.
   • Convert to File node to convert the B64 string to a file.
   • HTTP Request to ImageKit.io to host the image and get a URL.
   • Air Table node to create a new record in the "Generated Post" table with the post title, content, image URL, and content type.
     • Links the post to the original company.

Setting Up the Content Calendar

• Add a "Schedule Post" checkbox and a "Scheduled For" date field to the "Generated Post" table.
• Create a new page in Air Table Interfaces using the "Calendar" view.
• Select the "Generated Post" table.

Conclusion

The video provides a detailed walkthrough of building a productized AI system for LinkedIn content generation. It covers the key concepts, requirements, and steps involved in creating such a system, including setting up Air Table, n8n, and Air Table Interfaces, as well as using AI agents and image generators. The video emphasizes the importance of end-to-end workflow automation, user-friendly interfaces, and human-in-the-loop processes. While the system built in the video is an MVP, it provides a solid foundation for building more complex and valuable AI systems. The presenter also suggests improvements such as making the system more modular, allowing users to generate multiple images, and adding a review step.