AI Applications In Podcast Production

At Radio Milwaukee, we've always wanted to do more storytelling. But growing our capacity means hiring more producers, and that's not in the budget — especially with CPB funding cuts making things tighter across the industry. I decided to take action. I'm entering the DigitalOcean Gradient™ AI Hackathon, and instead of building another generic AI demo, I decided to build something for my industry. Something we actually need right now. It's called StoryForge — an AI-powered audio production platform designed specifically for public radio and podcasters. Here's the reality: we can't afford to hire more producers. But we already have incredible people in the building. Our DJs know their communities better than anyone — they're at community events, festivals, and concerts. They have the trust and the relationships. What they lack is training in narrative audio production or the tools to efficiently turn a great interview into a polished story. StoryForge changes that. It gives every DJ an NPR-trained AI editor in their pocket. How It Works We're using DigitalOcean Gradient™ AI to deploy four specialized AI agents that work like a production team: CoachAgent — Trained on NPR storytelling methodology. It doesn't just fix your script — it asks "What's the emotional core of this story?" and pushes you to find it. It coaches at every stage: story focus, tape selection, script writing, narration delivery, and even sound design. TranscriptAgent — Powered by Deepgram Nova-3. Turns raw audio into word-level transcripts with speaker identification. This is the foundation that enables text-based editing. ContentAgent — Takes one interview and generates six content formats: air break script, podcast segment, social thread, web article, newsletter, and press release. Each written in the station's voice AND the individual DJ's personal style. WorkflowAgent — Manages the full approval pipeline so nothing goes to air without a human producer signing off. Key Features - Text-based audio editing — delete a sentence in the transcript, the audio cuts automatically. No DAW required. - Real-time AI coaching at every production stage, trained on NPR Training principles - Auto-Assembly Engine — mixes narration, interview tape, music beds, and sound effects into a broadcast-ready piece - 5 mandatory human approval checkpoints — AI is a tool, humans are always in charge - One interview becomes six content formats automatically - Role-based permissions for Admins, Producers, DJs, Contributors Public radio doesn't have a talent problem. It has a capacity problem.

𝐀𝐈 𝐢𝐬𝐧’𝐭 𝐭𝐡𝐞 𝐟𝐮𝐭𝐮𝐫𝐞 𝐨𝐟 𝐰𝐨𝐫𝐤. 𝐈𝐭’𝐬 𝐚𝐥𝐫𝐞𝐚𝐝𝐲 𝐭𝐡𝐞 𝐜𝐨-𝐰𝐨𝐫𝐤𝐞𝐫 𝐰𝐡𝐨 𝐧𝐞𝐯𝐞𝐫 𝐬𝐥𝐞𝐞𝐩𝐬. I’ve stopped thinking of AI as “something to try” and started treating it like a trusted creative partner. The kind that never takes coffee breaks, doesn’t get offended when you tweak its work, and is always up for iteration number 9. Let me give you two very real examples: 🚀 At HackerRank | Scaling GTM with fewer people, more velocity In a world where your next buyer could be in San Francisco, Sydney, or Stuttgart, we needed to build a content engine that could adapt fast, localize fast, and ship faster. Here’s how we’re integrating AI across our marketing stack: HeyGen helps create avatar-led, studio-quality videos created entirely from text. We control tone, gestures, and language, which means we can go from idea to impact without booking a studio. Runway ML brings our animations and storytelling to life with movie-like quality. What used to take a full creative brief, a production agency, and 4–8 weeks to execute, now takes a few working sessions and a couple of renders. ElevenLabs enables us to explore voiceovers without needing studio time or VO artists. It’s early days, but we’re seeing massive potential in automating narration and creating AI generated video with real human audio. 𝐖𝐡𝐚𝐭’𝐬 𝐢𝐦𝐩𝐨𝐫𝐭𝐚𝐧𝐭 𝐡𝐞𝐫𝐞 𝐢𝐬 𝐭𝐡𝐢𝐬: None of these tools replace the people behind our content. They remove grunt work, spark creativity, and most importantly, give us back time, which is the real premium today. 🎙️ 𝐅𝐨𝐫 𝐦𝐲 𝐩𝐨𝐝𝐜𝐚𝐬𝐭, The Great Indian Points And Miles Show — 𝐂𝐫𝐞𝐚𝐭𝐢𝐯𝐢𝐭𝐲 𝐚𝐭 𝐒𝐜𝐚𝐥𝐞 When you’re running a passion project with a full-time brain on the weekends and a part-time team, AI is the enabler that bridges the gap. We use Suno to create original audio tracks for segments. No more digging through royalty-free libraries that sound like elevator music. Lovable helps us spin up event landing pages for community meetups in minutes. No dev dependencies, no bottlenecks, just fast GTM. ChatGPT sits in our ideation process: scripting intro hooks, breaking down credit card rewards jargon into human language, and yes, even helping me title an episode or two when I’ve got decision fatigue. These tools aren’t some shiny “tech stack” I’m flaunting. They’re behind-the-scenes partners helping me build something that’s real, resonates, and scales. 𝐖𝐡𝐚𝐭 𝐈’𝐯𝐞 𝐥𝐞𝐚𝐫𝐧𝐞𝐝: 1) AI isn’t here to replace people. It’s here to support great teams in doing even better work. 2) The magic isn’t in knowing the tools. It’s in knowing when to use them and how much to trust them. If you’re a marketer, a creator, or a curious tinkerer, don’t wait for a “perfect AI use case.” Start where the friction is. Chances are, your next breakthrough isn’t a brainstorm away, it’s a prompt away! 😃

AI has completely changed how our agency creates hooks for podcast clips. Instead of staring at a blank screen struggling to think of the perfect opener, we now have AI generate 10 different hook options at once. Then we pick the hook that feels right for the audience and modify it to match the creator's authentic voice. The human touch remains essential. And the genius of this method isn't letting AI write the final hook, you're just using AI as a springboard for your own creativity. Furthermore, what I love most is how this approach saves mental energy for the parts of content creation where human judgment truly matters, like deciding which moments from a 2-hour podcast deserve to become clips in the first place. So I want you to try this: Next time you're stuck on a hook, ask AI for multiple options, then trust your gut on which one resonates best with your specific audience.

I built a research assistant to streamline my podcast preparation process. For each episode, I create a research brief with my insights, guest background, topic context, and potential questions. This involves researching the guest and their company, reviewing their podcasts, reading their blog posts, and diving into the discussion topic—quite a time-consuming and effort-intensive process. To save time, I built an agent to handle this work. The project also showcases how to design an event-driven AI architecture, decoupling AI workflows from the app stack, leveraging event streams for data sharing and orchestration, and incorporating real-time data. It's built with: ◆ OpenAI various versions of GPT and Whisper ◆ LangChain for prompt templates and LLM API abstraction ◆ Next.js by Vercel ◆ Kafka and Flink on Confluent Cloud for agent orchestration and stream processing ◆ Bootstrap and good ol' fashion hand coded CSS for styling Behind the scenes: 1. Create a podcast research bundle with the guest name, topic, and source URLs 2. The web app writes the research request to an application database 3. A source connector pulls the data into a Kafka topic and kick starts the agentic workflow 4. All URLs are processed, text is chunked, and embeddings created and synced to a vector database 5. Flink and GPT is used to pull potential questions from the source materials 6. A secondary agent compiles all the research material into a research brief I cover this in detail here: https://lnkd.in/gSSBuC3t You can checkout the code here: https://lnkd.in/gUpY-YgQ #llms #agenticai #kafka #flink #confluentcloud

Google AI Studio and Gemini are my go-to tools when it comes to transcribing podcast interviews and formatting them so they’re worthy of a blog post and newsletter. Here is my exact workflow, with the prompts I used for my most recent podcast. I upload the MP3 to Google AI Studio, which excels at handling audio files. My prompt: "The attached file is an interview for a podcast called A History of Marketing between Andrew Mitrak and Sergio Zyman, Chief Marketing Officer of Coca Cola. It is about the history of New Coke, the Cola Wars in the 1980s and early 1990s. Please generate a clean transcript and remove "um" and other filler words and accidentally repeated words but otherwise be as accurate as possible." Providing context in the prompt (names, topic) makes for much more accurate output. I review the transcript in Google Docs using its error-checking features. I then upload this version of the transcript to my YouTube video, which is a big improvement over its auto-subtitles. Next, I use the Gemini App. I attach a PDF of journalistic transcribing instructions and use this prompt followed by the full text of the transcript: "The following is an interview transcript. Please make edits to correct grammar and remove false starts, following the attached transcribing instructions. Please format this for a blog and add line breaks when speakers alternate. When there is a long answer, break it up as needed into separate paragraphs for readability. Put the names of speakers in front of their dialogue each time they speak and bold their names." This cleans up the text, adds formatting, and attributes dialogue. The output at this point looks a lot like a blog post! I export to Google Docs. A 30-minute interview will be about 10 pages. For SEO and scannability, I use this prompt: "Please suggest SEO-optimized headers to add to this blog. Make them descriptive of sections. Keep them short, but don't try to be cute. Make sure they improve scannability. Use H2 and H3 formats." This generates headers I insert into the blog. I rewrite and edit these, but AI saves a lot of time here with the first draft. Finally, I review the blog post while listening to the MP3. This lets me check both the transcript and the audio file for errors simultaneously. At 2X speed this process takes 15-30 minutes. This workflow with Google AI Studio and Gemini has streamlined my post-interview process. It's not just about saving time, it's about producing something I otherwise wouldn’t have made without the help of AI. I wouldn’t bother with transcripts if I had to do them manually, so now the interview is more accessible to audiences who prefer to read instead of listen or watch the interview. It’s also more discoverable, and a better overall experience for everybody. Hope this long-form, detailed post is useful to those learning to use AI tools. I'm continuing to make this process faster each time. Would appreciate any of your AI tips if you have them!

This weekend, I was supposed to be working on shorts/reels and marketing posts for Her Career Unplugged. Instead, as I was unsupervised, I built an AI podcast editor. While working on clips for the podcast, I started looking at tools for generating shorts and editing multi-camera footage. The tools are good, but most of them require uploading hours of raw video to someone else’s servers and doing all the editing inside their platform. As an engineer, that made me pause a bit. I don’t mind cloud tools (I love the cloud), but I do care about understanding where my data is going and having control over my workflow. So naturally, instead of finishing the shorts I was supposed to be making, I built DarkRoom. DarkRoom is a local-first, multi-camera podcast editor. You upload your pre-aligned camera files, it transcribes them locally, generates an AI edit decision list using Claude, lets you review and tweak the cuts in a browser UI, and then renders the final exports using FFmpeg. Nothing leaves your machine (without you knowing). Ironically, starting Her Career Unplugged is what inspired the whole experiment. If you're curious about how it works, I wrote up the architecture and the pipeline here: https://lnkd.in/eevqr4as

This year I’ve been on 100+ podcasts which account for more than 50% of all of our inbound. For every podcast, it’s important for me to get the most scale out of every asset that I can pull. For most people, that’s where the leverage stops. Episode drops → one or two promo assets → onto the next one. But if you’re doing that, you’re leaving a ton of scale on the table. Because every episode actually contains: -Multiple sharp, 30–60 second takes -Reusable stories and frameworks -Hooks that could live on LinkedIn, YouTube Shorts, your site, email… The real game isn’t “be on more podcasts.” It’s: get the right clips out of every appearance and put them back to work everywhere. That’s what I started doing with Goldcast's new Agentic Video Editor. I take all of my appearances and drop them straight into Goldcast. The Agentic Video Editor analyzed everything and surfaced the exact moments where I was really on a roll: -Clean, self-contained clips -Strong hooks and one-liners -Segments where I’m explaining a framework or telling a story Instead of me hunting through 60-minute recordings, it did the first draft of “here are the clips worth scaling.” From there, I turned those moments into a whole library of ready-to-publish clips in a fraction of the time it would normally take. The AI handled the grunt work: -Auto captions I could quickly tweak -Smart scene changes so clips feel dynamic, not like a Zoom rip -On-screen callouts for the big ideas or quotable lines -Suggested b-roll and background music to keep people watching The result? Every podcast appearance now turns into: 1️⃣ A set of short clips for social 2️⃣ Mid-length cuts for landing pages, nurture, or sales enablement 3️⃣ A searchable library of “my best answers” on core topics Same recordings. Way more surface area. Way longer shelf life. If you’ve spent the year doing podcasts and panels, Goldcast’s Agentic Video Editor is the fastest way I’ve found to: Turn scattered interviews into a clip library And turn that clip library into real distribution—across every channel your buyers actually hang out on. #GoldcastPartner

Google's NotebookLM has introduced a new customization tool that allows users to create personalized AI-hosted podcasts. Users can upload documents and generate audio discussions, now with the ability to tailor content by adding prompts. The update lets users specify topics, focus on particular document sections, or target different audiences. While results vary, the tool offers significant flexibility, creating podcasts for entertainment, study, or niche audiences. Despite occasional mixed outputs, it shows promise in generating personalized AI-driven content.

I recently tried out NotebookLM to convert one of my old keynote talks into a podcast-style conversation. This was my original talk: https://lnkd.in/g4KBFpc9 and here is the 7-minute podcast it generated: https://lnkd.in/gXB6EJA6 It took a few tries and there were some minor hallucinations, the results were still super impressive! Even in its experimental phase, you can clearly see the potential for enterprise use cases—from content repurposing to generating learning materials. This could become a real game-changer, not just for boosting productivity but also as an engaging learning resource. Definitely worth exploring for anyone looking to use AI day-to-day beyond just chatbots.

I used to think great storytelling took a big team. Turns out it takes a great system. Here's mine... 7 AI tools I use working at a startup in San Francisco 👇 OpusClip — turns long-form podcast episodes into short clips automatically. It figures out the best moments so I don't have to scrub through hours of footage. Does most of the heavy lifting on our shorts across every channel. Descript — my podcast editing home base. Studio sound, captions, transcripts, YouTube descriptions with timestamps, social drafts, blog starters. Claude — The tool I'd give up last. It knows my voice, my projects, my style. I use it for everything from storyboarding and caption writing to building full interview prep packages. Last week I asked it to research a guest, and it came back with a 30-question arc, an intro script, a one-page cheat sheet, and a Google Doc dropped in the right folder. I have a whole library of prompts I've built up over time that turn raw material into impactful content (lemme know if you wanna see it 😜) It doesn't just save me time; it makes the work better. Composio — the bridge between Claude and all my work apps. Web search, Slack, Gmail, Drive, Notion — Claude can read from and act across all of them through one connection. For guest research that means Claude can pull the public web AND cross-reference it against my notes in Drive. Way more powerful than either piece alone. Riverside — where I record my online podcast interviews. Audio quality is genuinely great and remote recording feels seamless. Plaud AI— a pocket recorder that comes everywhere with me. Summarizes every meeting, surfaces action items, keeps me from losing a single idea. If you're a marketer, journalist, or content person, this one's non-negotiable. Epidemic Sound — for music. Not AI, but too good to leave off. It's wild how much more one person can do with AI in the loop. Most days it feels like I'm shipping what used to take a team of ten. Keep in mind, none one of these replace how I think. They just remove the friction between an idea and the output. What's in your stack as a marketer? LMK! Always looking for new tools to try out 😇 #AItools #ContentCreation #Storytelling #StartupLife