How I Built an Edge AI Pipeline to Automate Podcast Notes

When I first started recording my own podcast, I genuinely thought the hardest part would be getting good audio. I spent weeks researching microphones, treating my closet with acoustic foam, and learning how to edit out background noise. I was wrong. The real bottleneck happened long after I hit "stop recording."

Staring at a blank document to write episode titles and show notes was paralyzing. I would spend two hours just trying to come up with a catchy title or summarize a 45-minute conversation. It drained the creative energy I needed for my next episode.

I realized I wasn't alone in this. Many indie creators and solo developers face the exact same wall. So I decided to build something.

The Hidden Cost of Poor Podcast Packaging

We often talk about the content of a podcast, but we rarely talk about its packaging. Your title, description, and show notes are the storefront for your episode. If they are weak, it doesn't matter how brilliant the audio is—nobody is going to click play.

Search engines and podcast directories rely heavily on text to understand what your episode is about. If your show notes are just a single sentence saying "In this episode, we talk about coding," you are missing out on valuable SEO opportunities.

The cost of ignoring this is high. Great episodes get zero downloads because the metadata is an afterthought. As a solo creator, context switching from "engaging host" to "SEO copywriter" is mentally exhausting. Every time you switch from creative mode to analytical mode, you lose momentum. It's like trying to write a novel while simultaneously acting as your own literary agent and marketing team. I needed a way to automate the packaging without losing the human touch.

Building an Edge AI Pipeline

To solve this, I started experimenting with Large Language Models (LLMs). But simply pasting a transcript into an AI and asking for "show notes" usually results in a messy, unstructured wall of text. The secret is in the architecture and the prompt engineering.

I decided to build PodCrisp using Cloudflare Workers. If you are new to edge computing, think of it like this: instead of your code running on a single server in one geographic location, it runs on a global network of servers. When a user requests a tool, the code runs on the server closest to them. This means almost zero latency and no awkward "cold start" delays.

For the AI inference, I use models deployed directly at the edge. This keeps the data processing fast and secure.

Let's look at how I structured the show notes generation pipeline. Here is a simplified version of the Cloudflare Worker that handles the heavy lifting:

// Show notes generation pipeline (Cloudflare Workers + AI) export default { async fetch(request, env) { const { transcript } = await request.json(); // Call AI model for structured show notes const response = await env.AI.run('@cf/meta/llama-3-8b-instruct', { messages: [ { role: 'system', content: `Convert podcast transcripts to show notes: ## Episode Summary (2-3 sentences) ## Key Takeaways (5-7 bullet points) ## Timestamps (estimate from content flow) ## Resources Mentioned` }, { role: 'user', content: transcript } ], max_tokens: 800, }); return new Response(JSON.stringify(response), { headers: { 'Content-Type': 'application/json' }, }); }, }; 

Let's break down what is happening here for those who might be new to AI APIs.

First, we extract the raw transcript from the incoming request. Then, we call the AI model using env.AI.run. The magic happens in the messages array.

We use a system prompt to set strict rules. Instead of just saying "write notes," we explicitly define the markdown structure we want: a short summary, bullet points for takeaways, estimated timestamps, and resources. This structured approach drastically reduces hallucinations and keeps the output consistent. The user prompt simply passes the raw transcript. Finally, we cap the max_tokens to ensure the model doesn't ramble, and return the structured JSON.

For the title generator, I use a similar approach but with a different system prompt. I instruct the model to use curiosity gaps, avoid clickbait, and provide a numbered list. By separating the title generation from the show notes generation, each tool can be optimized for its specific task.

Practical Results and Lessons Learned

Building this pipeline taught me a lot about the realities of working with AI in production.

What worked incredibly well was the edge deployment. Because the inference happens on Cloudflare's network, the tool feels instantaneous to the user. There is no waiting around for a distant server to process the text.

I was also surprised by how much the system prompt mattered. I initially tried using a massive, highly complex prompt, but it actually confused the model. Keeping the instructions simple, direct, and structurally rigid yielded much better results.

If I were to change something today, I would refine how the model handles timestamps. LLMs are great at understanding text flow, but they don't actually "hear" the audio. Estimating exact timecodes purely from text can sometimes be inaccurate. In the future, I'd like to integrate a lightweight audio-sync tool to feed actual timecodes into the prompt, rather than relying on the AI to guess them.

Try the Free Title Generator

If you are struggling with the post-recording bottleneck, you don't have to stare at a blank page anymore. I packaged the core logic into a suite of tools, including a show notes generator, script outlines, and SEO keyword suggestions.

To start, I've made the Podcast Title Generator completely free to use. It uses the same edge AI principles to give you multiple title options based on your specific tone and topic. You can check it out and generate some titles for your next episode here: Podcast Title Generator.

I hope this helps you spend less time writing metadata and more time creating great content.

Full disclosure: I'm the developer of PodCrisp. I built this because I needed it myself.