Live Workshop: AI Pipelines & Agents in TypeScript w Mastra.ai - Nick Nisi and Zack Proser

On June 3rd, 2025, my colleague Nick Nisi and I hosted a two hour workshop at the AI Engineer World Fair in San Francisco. The event gathered more than 70 engineers eager to build AI powered workflows in their own products using Mastra.ai - a new framework that allows you to orchestrate AI tasks using pure TypeScript.

Live Demo: AI Meme Generator in Action

Here's our complete workshop demonstration showing the agentic application we built running end-to-end. Watch how workplace frustrations get transformed into shareable memes through a multi-step AI workflow:

How the AI Workflow Works

Users simply enter their random workplace frustrations - anything from "my meetings always run over" to "the deployment process is completely broken." Then our 4-step agentic workflow takes over:

1. 🧠 Analyzes their frustrations - Uses structured AI generation to extract and categorize the workplace pain points
2. 🎭 Selects the right base meme - Searches through Imgflip's extensive template database to find the perfect format
3. ✍️ Humorously recaptions it - Generates contextually funny text that transforms frustration into humor
4. 🚀 Creates and hosts it - Produces the final shareable meme with a stable URL

The result? Common workplace struggles become genuinely funny, shareable content that teams can actually use.

Key Learning Outcomes

By the end of the workshop, attendees had:

• ✅ Built a complete agentic workflow with multiple specialized steps
• ✅ Integrated external APIs (Imgflip, OpenAI)
• ✅ Used structured generation for reliable data extraction
• ✅ Created a fun, shareable application that solves real problems
• ✅ Learned Mastra's agent and workflow patterns
• ✅ Gained experience with AI-powered workflows and step composition

From Toy Example to Production Patterns

While we chose a humorous implementation for our workshop, the patterns, architecture, and Mastra framework we demonstrated are ideal for creating enterprise-grade agentic applications. The same workflow concepts apply whether you're building:

• Customer Support Automation - Multi-step ticket resolution workflows
• Data Processing Pipelines - ETL workflows with AI-powered transformations
• Content Generation Systems - Marketing copy, documentation, and reports
• Complex Business Logic - Approval processes and intelligent routing

The modular, typed, and testable approach we used for meme generation scales directly to serious production use cases.

The Meme Generation Pipeline: Technical Deep Dive

Our workshop demonstrates a complete 4-step agentic workflow that transforms workplace frustrations into shareable memes:

Understanding the Workflow Architecture

// Our main agent orchestrates the entire meme generation process
const memeGeneratorAgent = new Agent({
  name: 'MemeGenerator',
  instructions: 'Help users turn their work frustrations into funny memes',
  workflows: {
    'meme-generation': memeGenerationWorkflow,
  },
});

The workflow consists of four modular, reusable steps:

1. Extract Frustrations - Parse user input and categorize workplace pain points
2. Find Base Meme - Select appropriate meme templates from Imgflip's API
3. Generate Captions - Create contextually funny text using AI
4. Generate Meme - Produce the final shareable meme image

Step 1: Extract Frustrations with Structured Generation

One of Mastra's most powerful features is its integration with structured generation using Zod schemas and the Vercel AI SDK:

const extractFrustrationsStep = createStep({
  id: 'extract-frustrations',
  description: 'Extract and categorize user frustrations from raw input',
  inputSchema: z.object({
    userInput: z.string().describe('Raw user input about work frustrations'),
  }),
  execute: async ({ inputData }) => {
    const result = await generateObject({
      model: openai('gpt-4'),
      schema: frustrationsSchema,
      prompt: `Extract frustrations from: ${inputData.userInput}`,
    });
    return result.object;
  },
});

The frustrationsSchema ensures reliable data extraction:

const frustrationsSchema = z.object({
  primaryFrustration: z.string().describe('The main workplace frustration'),
  category: z.enum(['meetings', 'communication', 'process', 'technical', 'management']),
  intensity: z.number().min(1).max(10).describe('Frustration level from 1-10'),
  context: z.string().describe('Additional context about the situation')
});

Step 2: Find Base Meme Templates

The second step searches Imgflip's extensive meme template database:

const findBaseMemeStep = createStep({
  id: 'find-base-meme',
  description: "Get a diverse selection of meme templates from Imgflip's free API",
  inputSchema: frustrationsSchema.extend({
    analysis: z.object({
      message: z.string(),
    }),
  }),
  execute: async ({ inputData }) => {
    const response = await fetch('https://api.imgflip.com/get_memes');
    const data = await response.json();
    
    // Select templates that work well for workplace humor
    const selectedMemes = data.data.memes.slice(0, 10);
    return { templates: selectedMemes };
  },
});

Step 3: Generate Contextual Captions

This step uses AI to create relevant, funny captions based on the extracted frustrations:

const generateCaptionsStep = createStep({
  id: 'generate-captions',
  description: 'Generate meme captions based on frustrations and template',
  inputSchema: z.object({
    frustrations: frustrationsSchema,
    baseTemplate: memeTemplateSchema,
  }),
  execute: async ({ inputData }) => {
    const captions = await generateObject({
      model: openai('gpt-4'),
      schema: captionsSchema,
      prompt: `Create funny captions for ${inputData.baseTemplate.name} 
                     based on: ${inputData.frustrations}`,
    });
    return captions.object;
  },
});

Step 4: Generate the Final Meme

The final step uses Imgflip's caption API to create the actual meme:

const generateMemeStep = createStep({
  id: 'generate-meme',
  description: "Create a meme using Imgflip's caption API",
  inputSchema: z.object({
    baseTemplate: memeTemplateSchema,
    captions: captionsSchema,
  }),
  execute: async ({ inputData }) => {
    const formData = new URLSearchParams();
    formData.append('template_id', inputData.baseTemplate.id);
    formData.append('username', process.env.IMGFLIP_USERNAME || 'imgflip_hubot');
    formData.append('password', process.env.IMGFLIP_PASSWORD || 'imgflip_hubot');
    formData.append('text0', inputData.captions.topText);
    formData.append('text1', inputData.captions.bottomText);

    const response = await fetch('https://api.imgflip.com/caption_image', {
      method: 'POST',
      headers: { 'Content-Type': 'application/x-www-form-urlencoded' },
      body: formData,
    });
    
    const result = await response.json();
    return { 
      imageUrl: result.data.url, 
      pageUrl: result.data.page_url 
    };
  },
});

Real Examples from Our Workshop

Here are actual memes generated during our workshop, showcasing how workplace frustrations get transformed into shareable humor:

Open Office Nightmares → Workplace Reality

When they move all the engineers to an open floor plan with noisy co-workers...

Compensation Blues → The Talent Paradox

Being told you're excellent at your job but somehow still underpaid...

The Helper's Dilemma → Salary Irony

When you refer an engineer who constantly asks for your help but gets paid more than you...

That awkward moment when your mentee's starting salary is higher than your current one...

These examples demonstrate how our AI workflow transforms common workplace frustrations into contextually appropriate and genuinely funny memes that teams can actually share!

Why Mastra.ai for Production Workflows

Modular Step Design

Each workflow step is designed to be:

• Modular: Can be used independently or as part of larger workflows
• Typed: Uses Zod schemas for input/output validation
• Testable: Can be tested in isolation using Mastra's built-in tools
• Reusable: Can be composed into different workflows

Built-in Testing and Development

Mastra provides excellent development experience:

# Start the development server
npm run dev

# Visit the Mastra playground at http://localhost:4111
# Use the chat interface to test the complete workflow

Error Handling and Reliability

The framework includes graceful error handling and recovery mechanisms, essential for production AI applications.

Workshop Experience and Results

All of the materials are open source in the mastra-agents-meme-generator repo. We built the full app first, wrote up the steps in workshop.md, and staged each phase on a separate git branch so everyone could work at their own pace.

The hands-on format let attendees build a simple pipeline from scratch, then wrap it with an agent interface. Seeing the agents successfully complete tasks through our pipelines was an awesome moment for everyone in the room.

One of us stayed on the mic while the other roamed the room answering questions and fixing issues. This tag-team approach kept everyone moving forward and made sure no one fell behind.

We tied the workshop back to lessons from my earlier talk at a16z about taking prototypes to production. This time, it was all coded live in TypeScript with the Mastra.ai SDK.

By the end, almost every attendee succeeded in getting the finished agent running. The energy in the room was incredible, and we were fired up by how many developers were brainstorming ways to use Mastra.ai in their own workflows.

Glowing Feedback from Attendees

The response from workshop attendees was overwhelmingly positive. Here's what engineers said about the experience:

The feedback validated our approach of combining humor with serious technical patterns - attendees appreciated both the engaging format and the practical, production-ready skills they gained.

Beyond Memes: Enterprise Applications

While our workshop focused on meme generation, the same patterns apply to serious business applications:

Workflow Types

• Customer Support: Multi-step ticket resolution workflows
• Data Processing: ETL pipelines with AI-powered transformations
• Content Generation: Marketing copy, documentation, reports
• Business Logic: Complex approval and routing processes

---

Interested in AI Workshops for Your Team?

Want to bring hands-on AI training to your organization? I provide workshops and training sessions on building production AI workflows, agentic systems, and practical machine learning implementations.

We can't wait to run more of these sessions in the future. If you want to follow along with the workshop materials yourself, head to the mastra-agents-meme-generator repo.