The Problem Every Educator Faces

Let’s be honest – marking assignments is a pain in the arse. I’ve got mates who are teachers, and they spend their weekends buried under piles of papers, squinting at handwriting that looks like it was done during an earthquake. And that’s just the text! What about diagrams, arrows pointing to different parts of drawings, crossed-out text that may or may not be intentional?

This client came to me with exactly this headache. They had course materials, assignment rubrics, marking guides, and heaps of other documentation scattered across Google Drive. Students were submitting handwritten assignments (yes, in 2025!), and the marking process was taking forever.

Enter the RAG Database Builder

If you haven’t seen my previous video about the RAG database builder, you might want to check it out first because that’s the foundation of everything I’m showing you today. Basically, this workflow grabs everything from your Google Drive – PDFs, Word docs, course manuals, rubrics, the lot – and throws it all into a searchable database that AI can query.

Watch the video and get the RAG Builder Workflow here.

The beauty is that once you’ve got this knowledge base set up, you can build AI agents that actually know your course content inside and out. They’re not just making stuff up – they’re referencing your actual materials.

The Assignment Marking Workflow

The workflow I built has two main parts, and both are pretty bloody impressive if I do say so myself.

Part 1: The Assignment Marker

This is where it gets interesting. The system connects to a Google Drive folder where students submit their assignments. Now, these aren’t neat typed documents – we’re talking handwritten assignments with cursive text, crossouts, diagrams with arrows pointing all over the place, and drawings that would make a kindergartner proud.

Here’s what happens when the workflow kicks off:

The system grabs the assignment file and sends it to Gemini document vision. This isn’t just text extraction – it’s actually reading the document like a human would. It understands that a line through text means it’s crossed out, it can follow arrows pointing to different parts of a diagram, and it can read cursive handwriting better than some humans I know.

Then it passes everything to an AI agent that’s been trained on all the course materials. This agent doesn’t just mark the assignment – it provides detailed feedback including the student’s answer, the mark out of the total possible, the reasoning behind the mark, and here’s the kicker – it references the exact text from the course materials that supports its decision, complete with document titles and page numbers.

The results get output as both a Google Doc and a styled HTML file. The HTML version looks much nicer with colour coding and better formatting, making it easy to scan through the results.

Part 2: The Interactive Q&A System

But wait, there’s more! (I sound like a bloody infomercial.) The second part of the system lets you have a proper conversation with your course database. You can ask questions like “Can you describe the intonation pattern in the English language diagram?” and it’ll not only give you a detailed answer but tell you exactly which document to find it in and what page number.

I tested it by asking about specific diagrams, and the AI could describe complex visual elements in detail. It understood flow charts, could follow the progression of information, and explained how different elements connected to each other. When I asked where to find test answer documents, it gave me direct links to the relevant files.

Real-World Results That’ll Make You Do a Double-Take

During the demo, I watched this thing mark an assignment that included:

• Cursive handwriting (the kind that gives humans headaches)
• Multiple crossouts and corrections
• Complex diagrams with annotation lines
• Technical terminology specific to the course

The AI correctly identified when a student had written “respect for the other person’s model of the world” in cursive, understood which parts were crossed out, and even analysed diagrams showing eye movement patterns with labels like “VC top left” and “AC middle left.”

For one question worth 6 marks, it gave the student full marks and explained: “The student correctly identified the eye patterns because in the image, the answer’s top left is VC, etc.” It actually understood the spatial relationships in the diagram!

When students got partial marks, the system explained exactly what they got right and what they missed. For example, “The student correctly identified 5 of 7 items” and then provided the complete correct answer for reference.

The Technical Bit (For Those Who Care)

The system uses n8n workflows to orchestrate everything. The RAG database is built using vector embeddings, which means it can find relevant information even when you don’t use the exact same words. The AI agents are prompted to always reference source materials, provide reasoning, and maintain consistency with the marking rubrics.

The beauty of this setup is that it’s not just a black box spitting out marks. Every mark comes with justification, source references, and clear explanations. If a teacher wants to double-check something, they can click directly to the relevant page in the course materials.

Why This Matters (Beyond Just Being Cool)

Look, I’m not trying to put teachers out of work. What this does is handle the grunt work of marking so educators can focus on the stuff that actually matters – like helping students understand concepts, providing personalised guidance, and developing better course materials.

The consistency is also a game-changer. Human markers have bad days, get tired, and sometimes mark differently depending on whether they’ve had their coffee. This system applies the same standards every single time.

Plus, the speed is just insane. What might take a human marker hours to complete, this system does in minutes. And it’s available 24/7, so there’s no waiting around for results.

Setting This Up for Your Own Use

If you’re an educator or work with educational institutions, this kind of system could be a proper game-changer. The RAG database builder I mentioned earlier is the starting point – you’d need to get all your course materials into a structured format first.

The assignment marking workflows would need to be customised for your specific marking rubrics and requirements. Each course is different, so the AI agents need training on what to look for and how to apply your marking criteria.

The good news is that once it’s set up, adding new courses or updating materials is pretty straightforward. The system can handle multiple courses simultaneously, and you can keep refining the marking criteria as you go.

Final Thoughts

I’ve been building AI workflows for a while now, but this one genuinely impressed me. Watching an AI system read handwritten text, understand complex diagrams, and provide detailed, referenced feedback feels like living in the future.

The fact that it can have intelligent conversations about course content and point you to specific pages in specific documents is just the cherry on top. It’s like having a teaching assistant that’s read every single course document and remembers exactly where everything is.

If you’re in education and dealing with the marking grind, this might be worth exploring. The technology is there, it works right now, and it’s only going to get better.

What do you reckon? Are you ready for AI to start marking your assignments, or does the thought terrify you? Let me know in the comments – I’d love to hear from educators who’ve tried similar systems or are thinking about it.

And if you want to build something like this yourself, grab the RAG workflow and start experimenting. The future of education is here, and it’s pretty bloody exciting!

Want to build your own AI workflows? Check out my other tutorials and grab the resources you need to get started. And remember – if I can figure this stuff out, so can you!

The post How I Built an AI That Marks Handwritten Exams in Minutes (And It’s Scary Good!) appeared first on SammyAI.

Ever tried uploading a 50MB PDF to an AI service or automation tool, only to hit a file size error?

This guide shows you how to bypass those limits with a free, open-source workflow that chunks massive PDF files into smaller, manageable parts — perfect for feeding into AI models, parsing tools, or cloud storage.

Built on top of N8N and a self-hosted PDF chunking microservice, this workflow is fast, flexible, and fully customisable. Whether you’re prepping files for Gemini, Claude, ChatGPT, OCR, or even internal review — this tool makes it possible.

Why You Need PDF Chunking

Most automation tools and AI APIs choke on large PDFs due to:

• Max upload sizes (often 5MB–20MB)
• Page count restrictions
• Timeouts or memory constraints

This setup eliminates those limits by:

• Breaking down PDFs into smaller chunks (by size or page count)
• Returning download links for each chunk
• Allowing each chunk to be processed independently in any workflow

What You’ll Get

• N8N workflow file (drag & drop ready)
• Custom PDF chunking microservice using pdf-lib
• Setup guide for Ubuntu (works with any VPS)
• Gemini prompt example (optional, not required)
• Support for large PDF files

Real-World Use Cases

• Feed chunks into AI tools like Gemini, OpenAI or Anthropic
• Upload to cloud storage or backup systems without hitting file limits
• Pass through OCR or document extraction services in batches
• Split course materials or books into manageable parts
• Pre-process for searchable indexing or embeddings

Gemini-Compatible, but Not Locked In

The example workflow includes a Gemini Document Vision fo analysing each PDF chunk, but you can swap these nodes to do whatever you want your PDF chunks.

Get Started

1. Download the workflow and install guide
2. Set up the chunking service on your server (10 mins)
3. Connect your PDF input source (Google Drive, upload, etc.)
4. Use the chunked files however you want!

Free, Open, and Yours to Control

Unlike paid APIs that gate your usage, this solution runs on your own terms, with:

• No file size limits
• No rate caps
• No hidden costs

Just host it, and start processing.

The post 🚀 Free Alternative to Paid PDF APIs – How to Chunk Large PDFs for AI or Automation appeared first on SammyAI.

I was building a RAG system for an educational client and…..their handwritten PDF assignments had students circling diagrams and standard PDF extraction wasn’t cutting it so I turned to Google Gemini’s vision API (so it can analyse entire documents including all the visual stuff, hand written text…you name it!).

Thought I’d share this as I see a few others have had this similar issue.

Key Steps:

Replace PDF Extract with Base64 Conversion – Instead of using standard PDF extraction, convert your PDF to Base64 format using n8n’s conversion nodes

Set Up Gemini API Integration – Create an HTTP Request node with POST method pointing to Google’s Gemini API endpoint for document analysis

Configure Your Prompt – Use prompts like “Give me the complete text as written, and when you encounter diagrams or images, describe them in detail including any annotations and how they reference the content”

Handle Processing Time – Gemini vision takes significantly longer than text extraction, so plan your workflows accordingly and consider breaking large documents into smaller chunks

Choose the Right Use Case – This method is perfect for educational assignments, handwritten documents, or PDFs with important diagrams, but stick to standard PDF extraction for large text-only documents

Expect Detailed Output – Gemini often produces more descriptive text than the original document, providing comprehensive analysis of visual elements and their spatial relationships

Setup Instructions:

Get Your Gemini API Key:

1. Go to aistudio.google.com
2. Navigate to API keys section
3. Create and copy your API key

Add Header Authentication in n8n:

1. Create new credential type: “HTTP Header Auth”
2. Header name: x-goog-api-key
3. Header value: <Your copied API key
4. Apply this credential to your HTTP Request node

The post 💫Converting PDFs with Images and Diagrams Using Google Gemini Vision in n8n appeared first on SammyAI.

I’ve just finished my own N8N migration recently from one server to another so thought I’d document and video it in case any of you ever consider doing this yourselves.

In this guide, I’ll walk you through every step—including setting up your new host, backing up all your N8N data, securely copying everything across, and getting your workflows up and running on the new machine.

For the demo, I actually migrate N8N from Digital Ocean to Hetzner. I’ve found Hetzner offers really affordable hosting options that are well-suited for running N8N and other Docker-based services. (Just to clarify: I’m not affiliated with Hetzner or earning anything from this—it’s purely a tip for anyone looking to save on hosting costs.)

What’s covered in the tutorial:

• Setting up your new host
• Backing up all your N8N data on your current host
• Copying all the data to your new host
• Restoring your N8N instance and getting it running again

I hope this helps anyone planning a migration or just curious about the process. Check out the attached document and video, and let me know if you have any questions!

Download Doc

The post 🚀How to Migrate N8N (Docker) to a New Host: Step-by-Step Guide + Video appeared first on SammyAI.

Thought I would share a workflow I’ve built for branded product image generation that uses Fal.AI, Cloudinary, and Google Sheets for a fast, scalable solution.

This setup leverages Fal.AI’s Flux LoRA training to create high-quality, consistent images featuring your models (people), products, and branding elements (logos/overlays). It’s particularly useful for building a content pipeline for ecommerce, social media, or promotional material.

After experimenting with other tools like Replicate, I found Fal.AI to be significantly faster, with a cleaner UI and a more intuitive experience — especially if you’re testing via their API Playground.

Workflow Overview

The system integrates the following tools:

• Google Sheets – Used as the control panel for:
• Google Drive – For uploading and storing training image files
• Catbox.moe – Used to publicly host zipped training image sets
• Fal.AI – For both LoRA model training and Flux image generation
• Cloudinary – For overlaying branded elements (logos, badges, etc.) on top of the generated image

Use Case

This workflow is ideal for generating consistent, on-brand product shots featuring your model(s) with minimal manual input. Great for ecommerce stores, marketing teams, or agencies handling multiple client brands.

The Download Includes:

• The JSON workflow file
• A notes document with links to API docs
• Google Sheets Template

Playing with Heygen for the video too

The post 🔥Branded Product Image Automation Using Fal.AI, Cloudinary & Google Sheets appeared first on SammyAI.

Turn ANY Google Drive folder into an AI-powered knowledge base in minutes!

What if I told you could transform your entire Google Drive into a searchable AI brain that rivals enterprise RAG systems?

THE WOW FACTOR

Watch this: Point it at ANY Google Drive folder → It automatically:

• Recursively scans ALL subfolders and files
• Intelligently processes PDFs, Google Docs, Sheets, even IMAGES
• Creates vector embeddings using OpenAI
• Auto-categorizes everything based on folder structure
• Detects changes and only updates what’s new/modified
• Serves up an AI agent that can answer questions from your data

Result? A fully searchable AI knowledge base that knows EXACTLY where every piece of information came from!

WHAT MAKES THIS SPECIAL

Plug & Play Revolution

• Zero configuration headaches – just update 4 variables and you’re LIVE
• Works with ANY Google Drive structure – your folders become categories automatically
• Drop into any workflow – designed as a reusable sub-workflow

Architecture

• PostgreSQL + pgvector for lightning-fast similarity searches
• Intelligent change detection – only processes what’s actually changed
• Automatic cleanup – removes deleted files from your database
• Folder rename handling – maintains relationships when you reorganize

AI-Powered Content Processing

• Multi-format support: PDFs, Google Docs, Sheets, Images
• GPT-4 Vision extracts text from images and screenshots
• Smart Excel processing converts tables into searchable text
• Maintains source attribution for every piece of information

Hierarchical Organization System

• 4-level categorization automatically built from folder structure
• Parent categories let you connect multiple workflows
• Relationship mapping between documents and categories
• Metadata-rich searches with filtering capabilities

REAL-WORLD USE CASES

Educational Institutions: Transform course materials into AI tutors
Corporate Knowledge: Make company documents instantly searchable
Research Teams: Turn research papers into queryable databases
Consultancies: Create client-specific knowledge bases
Healthcare: Organize medical documents with AI-powered retrieval

GET STARTED IN 5 MINUTES

1. Import the workflow (JSON attached)
2. Update 4 variables in “Set Globals”:
   • Google Drive folder ID
   • Database table names
   • Parent category name
3. Run the setup nodes (creates tables automatically)
4. Execute and watch the magic happen!

The post 🚀 GAME-CHANGER: Complete Google Drive to Vector Database RAG Workflow appeared first on SammyAI.

So here’s a simple form builder designed specifically for N8N webhooks that builds your own self-hosted forms. Nothing too flashy, it just lets you create a form, hook it up to your N8N instance, and download it as a single HTML file that you can run anywhere (even locally). All forms are built to support the webhook response data in N8N and I’ve done my best to format different JSON outputs in a readable format. It also has a voice recording field which was one of the features I use most…so I had to include that

Using the app is pretty simple, here it is in nutshell.

– Add your N8N webhook URL in the settings

– Choose from a few simple styles

– Add any fields you want (including a voice recorder)

– Set up conditional logic if you need to show/hide fields based on answers

– Export you form and go live!

Here’s a link to the tool https://formbuilder.idm2.com/

It’s been super handy for my own workflows, so I thought I’d share it in case anyone else finds it useful too. I’ve recorded a quick video walkthrough if you’re curious to see how it works.

Open to feedback or suggestions…keen to hear if this could help in your own projects.

Hope you’re all having a great weekend!

The post 💡N8N Webhook Form Builder appeared first on SammyAI.

What sets this workflow apart is its holistic approach to automation. It features real-time competitor monitoring, smart voice automation, and product-focused visuals using state-of-the-art image editing with tools like Fal’s Flux LoRA and Cloudinary overlays. Need to launch a campaign quickly? This workflow doesn’t just create the content, it automatically generates hero images, product images, and even promotional videos to accompany your newsletters. Plus, with built-in GitHub Gist previews and a man-in-the-loop approval process, your team remains in full control while saving hours on content creation and coordination.

Ready to elevate your marketing with smart automation and eye-catching visuals? Check out the video tutorial to see the full power of this workflow in action, complete with dynamic image and video examples, seamless newsletter creation, and streamlined campaign scheduling that’s as effortless as it is effective. Let’s unlock the next level of content marketing for your business.

Brand Image Generation using Flux LoRA Training – Watch Video

The post 🔥Supercharge Your Marketing with Our All-in-One AI Workflow: Content, Images, and Video – Fully Automated appeared first on SammyAI.

Key Features

• Full “Man-in-the-Loop” Approval: Every campaign requires review, you can Approve, Cancel, or Request Changes before anything is sent.
• Dynamic Newsletter Workflow: Rotates between general, featured product, and coupon/offer categories for variety (no more stale templates). It works on a 4 week (monthly) rotation with a different newsletter sent out each week.
• Image Generation: Uses Fal’s Flux LoRA for training and generating unique campaign hero images, always on-brand.
• Product Data Scraping: Firecrawl API scrapes Stores product catalog to auto-select and feature different products in campaigns.
• Auto Coupon Code Generation: Instantly creates new, limited-time coupons in WooCommerce for each relevant campaign.
• HTML template creation: HTML template published as Gist to Git for previewing and to Klavio when template is approved.
• GitHub Gist Previews: Newsletter drafts are auto-published as Gist for fast preview, review, and team feedback before approval.

Technical Overview

Firecrawl API:

• Scrapes live product data (name, image, price, link) from WooCommerce.
• Enables random product selection for featured-product mailouts.

Fal.AI API (Flux LoRA):

• Custom-trained image models using your brand’s assets.
• Generates fresh, unique images for every campaign—no repeats.

WooCommerce API:

• Creates single-use coupon codes with expiration (e.g., 2 weeks).
• Syncs codes straight into your online store for instant redemption.

GitHub API (Gist):

• Renders a newsletter preview as a Gist for collaborative review (sent in man in the loop email)

Man in the Loop Approval:

• Email sent asking to Approve, Reject or Request changes
• Change request are sent back AI Agent with memory of previous output and change request

Klaviyo Integration:

• Generate and publish a HTML template
• Sets up and schedules a campaign and associates the template to the campaign

Use Cases

• Completely Hands Free Mode. Set-and-forget weekly campaigns, always fresh, never missed.
• Hybrid Mode. You can continue manually running campaigns and approve workflows only when short on time or if they forgot
• Creativity Mode. You can use it to spark some creativity. Use the ‘Try again’ option in man in the loop to generate alternat versions.

The post 📨 Klaviyo on AutoPilot With N8N appeared first on SammyAI.