Daniel Mishler, Head of Automation at Deepgram. Deepgram is Voice AI platform for enterprise use cases – speech-to-text, text-to-speech, and speech-to-speech APIs for developers.

Listen to the episode on Spotify, Apple Podcast, Podcast addicts, Castbox. You can also watch this episode on YouTube.

In this episode of the AI Minds Podcast, Daniel Mishler, Head of Automation at Deepgram, shares his unconventional journey from naval architecture to leading automation at one of the most innovative AI voice platforms.

Daniel discusses how a growth mindset and relentless curiosity guided his transition into tech, and how Deepgram fosters a culture that empowers employees to experiment, learn, and build with cutting-edge tools.

He dives into his philosophy on automation—not as a means of replacement, but as a path to workforce augmentation, improving both efficiency and employee satisfaction.

The conversation explores Daniel’s work with Vibe coding, a no-code/low-code approach to building with AI, and how tools like Cursor and Windsurf helped him go from idea to real, functioning applications.

Daniel also shares the story behind Vibe Coder, a tool he developed that enables voice-driven programming, and his vision for a future where voice interfaces unlock creative coding for everyone.

Throughout the episode, Daniel emphasizes experimentation, hands-on learning, and the transformative role AI can play in reshaping how we work, build, and collaborate.

Show Notes:

00:00 From Novice to Head of Automation

04:16 Automation as Human Augmentation

06:41 Automating Insights from Slack Data

12:47 Iterative Investment in Automation

14:02 Creating Effective PRDs for VC Investments

17:26 Vibe Coding with Voice Integration

22:27 "Future of Voice-Driven Computing"

25:02 "Cursor's Potential for Complex Apps"

More Quotes from Daniel:

Demetrios:

Welcome back to the AI Minds Podcast is a podcast where we explore the companies of tomorrow being built AI first. I'm your host as always, Demetrios and this episode, like every episode, is brought to you by Deepgram, the number one speech to text and text to speech API on the Internet today. Trusted by the world's top enterprises, conversational AI leaders, startups, and some of them you may have actually heard of like Spotify, Twilio, NASA and Citibank. I have the pleasure today of being joined by Dan, who is head of automation at Deepgram and recently came out with Vibe Coder. How you doing, Dan?

Daniel Mishler:

I'm doing well, thanks for having me.

Demetrios:

I'm excited to talk to you all about Vibe coding today. As you know, it might be the top buzzword of 2025. I think it's close competition between MCP and Vibe Coding. I have fully changed my title on LinkedIn to be a Vibe Manager because I manage the good Vibes only. But I want to know about you and your journey into tech. I know that you still started in product and just give me like a little background.

Daniel Mishler:

So went to Coast Guard Academy in college and thought I wanted to be a naval architect. Realized naval architecture jobs in Michigan where I wanted to live are kind of few and far between. So I had to pivot once I got out of the military and went and got my mba, pivoted fully into business and that thought that's what I wanted to do. Did management consulting for a while. We can talk about that's a fun topic. But post that really gave me my opportunity to dive into tech and I actually started on the hardware side where I joined and got the startup bug. Working for a small satellite design company that was designing synthetic aperture radar satellites.

Daniel Mishler:

And I knew nothing about synthetic aperture radar satellites, but I'm always eager to learn. I'm a big believer in the growth mindset, which we'll definitely talk about later when it comes to Vibe coding. But from there got the startup bug. Never wanted to go back to a big corporation. So one thing led to another and I was able to worm my way into Deepgram and started as a product manager, quickly shifted over to program manager because we needed some help on that front based on my prior experience and then got an opportunity to be the chief of staff for a year and now I've recently transitioned over to the head of automation role where my mission is to make Deepgram the most productive and efficient company in the world using AI tools and Kind of exploring the kind of the innovation that's coming out every day and the opportunities that's being presented there.

Demetrios:

I know from the inside how much the teams and the folks at Deepgram are creating workflows with AI. I know that you've got wild ideas on different things that we can do. Do you have a intuition on why that is, that we are so AI native per se?

Daniel Mishler:

I think Deepgram, first and foremost we hire fundamentally curious people. That's a big part of our hiring process is are people willing to learn and try the latest things on their own time, on company time doesn't matter like what is the most effective way to do the job. We hire people who have strong opinions loosely held and so if something new comes along that does the job better, generally speaking, we shift to it and we adopt it if we think it's valuable and there's no ego, there's no this is the way we do it. No, let's figure out the best way to do things and make our customers the happiest customers on the planet and we'll go from there.

Demetrios:

What are some automations that either you have implemented or you've seen other teams implementing that you've been really excited about?

Daniel Mishler:

I think when it comes to automations, it's hard to just sit down and be I want to automate my job away. Like no one wants to actually automate their job away first of all. So you have to be careful around the messaging there because the idea is not to replace humans, it's to augment humans. And so I almost, I kind of nitpick at the word automation really. I really view it as augmentation. How can we make ourselves more effective using these AI tools and remove some of the tedious, boring work and enhance our own abilities and skills with these tools. And in doing so now you're not only are you having a lot more fun at your job, which is 100% the case for me, but now you're also delivering a better experience for your customers and making your company more successful and at the same time, but specifically I think the biggest for us, when you sitting down and looking at it, there's just so much opportunity to automate, you can kind of get lost. And so you just got to start with the low hanging fruit.

Daniel Mishler:

There's low hanging fruit and then there's these big long shots and start with something small, get a quick win. And once you do that then you can Build a momentum. You can kind of prove to the organization this is something that's worth spending time on. And then from there you can start working on those long shots. Like I have this crazy idea that if we do this xyz, then here's the benefit. But it's going to take time to justify and maybe the tools aren't there yet, but they're heading that direction and we can get, proof of concept and stuff. But it's never a good idea to just sit down and only do that or only do low hanging fruit. You need both, I think to really show and kind of get to the potential of automation.

Daniel Mishler:

To be clear, we're just starting on a journey. Like we're already a very efficient company, but I think there's a lot that we can do and a lot of the, that we can tap into to become even more efficient and productive without, placing an additional burden on our employees.

Demetrios:

One that I've set up for myself is from these podcasts. How do we capture insights of Deepgram users and then turn it into a case study from what folks say on the podcast? And automating that journey has been my kind of like fun side project to create.

Daniel Mishler:

That's an amazing use case. When I sit down, I think about the amount of information that's just contained in Slack, Like you have this institutional knowledge that's in unstructured format within these Slack channels or within company meetings. You record everything, hook it all up to a vector database and start to just explore like what can you do? What are the different pipelines you can set up? We have an amazing team in our revenue operations side of things that has set up automations around external phone calls, plugging those transcripts using our technology, of course, dog food as much as possible, plugging that into an LLM based workflow that can then analyze each call, identify the product line that's associated with, identify challenges, risks, opportunities with each customer, get some great product feedback automatically and then summarize it all in posting in Slack and that's something that you used to have to do manually. You had to set aside time all right, I'm going to take an hour every day to look at the recordings and see and then write up a report. Now it's just done. And you can, there's downsides here too. let's be realistic, you can get information overload.

Daniel Mishler:

So you need a really effective way of triaging all those so that they just don't get dumped into a slack channel and then nobody reads them. And then nobody reads them and someone asks the question anyways on slack. so there a balance here. I mean, I'm bullish on AI and AI augmentation, but there's some realistic, barriers you have to overcome when it comes to this new way of working.

Demetrios:

Because I look at it as in best case scenarios, it's that the cost per unit of intelligence has dropped dramatically. And so if we consider a token a unit of intelligence, then that's a fancy way of saying it. And that's best case scenario. That's what we're looking at. And worst case scenario, some days I look at it and I say these damn LLMs are so disrespectful of my time. The amount of text they output just so that I can find this one little piece of actual information that I need is absurd.

Demetrios:

And so I think you do right to call out that you potentially can get this information overload or you can be constantly pinged if you're not designing the system in the right way to, to actually bring that value when you want it. And in a way that is digestible 100%.

Daniel Mishler:

I mean you see two camps, You get the automated everything. AI is going to take all the jobs away at camp and then you get the camp that's these models are garbage. And AI is useless and it's just a trick, It's a party trick. Like I think the answer is in between. again it all comes down to can you exactly what you said really guide the LLM to respond in the proper way and get the true value out of it. Being careful of hallucinations, being careful of confirmation bias and it just telling you what you want to hear. So careful prompting is key, careful pipeline setup is key. But the big thing is, if you're not culturally willing to try it and just get started and try it, and you're not kind of building that in your company, you're going to be behind the eight ball because it is going to happen.

Daniel Mishler:

The models are getting better. And so you don't want to be left behind just because you're a naysayer. But at the same time you do have to, you run the risk of turning off the wider culture of your company if you say, here's this amazing automation that I built, it's perfect, it does Everything just right and then it burns someone. Now they're permanently turned off the idea of automation and they get pushed into the latter camp of AI is useless. So it is a careful balance to walk, making sure, this is the future. We're on the uptick though, There's going to be the valley of disillusionment soon. Maybe we're entering it now, I don't know in some regards, but when all the flaws get revealed, but eventually it's going to peak back up, This is when the PC came out, people said, this is stupid. Like no one's going to use it.

Daniel Mishler:

When the Internet came out, same thing, when the mobile, when the iPhone came out. It happens every time. So this is the most predictable thing in the world. Push through it, identify the flaws, be real with people about, we're early adopters, but we need to do this now so that as this technology matures and it's maturing faster than any other kind of game breaking technology has ever matured, ever, we can be on the forefront of it and really catch the wave at its best perk.

Demetrios:

I like something you said there with respect to almost like we have to build the muscle and maybe you put 10 workflows into the company and nine of them are not good, but that one could be absolute fire that gives the so much time back to your workforce and you wouldn't know that one. It's just like when you're testing out products and you're trying to go to market and you're figuring out do we have product market fit here? It's that same idea. You got to kind of throw spaghetti at the wall, see if you like it, see if it works, if it doesn't take more time from you than you actually are putting into it, or it's not asking you to read a whole report every other call that the whole company is on. And so you're like, oh man, the, you, you have to think through these things, but you don't know that until you try it.

Daniel Mishler:

100% agree on that one. Because you run the risk of being scared or you run the risk of being bullish, Too bullish. And then you get in trouble real quickly. And I think what I would recommend and what I've been thinking about a lot is taking like the VC approach here where you build quickly, you make lots of little investments, And with the hope that one really takes off and like that's the one that you leverage then to then make more investments. And it's the same thing with building these automations is make a bunch of investments, small ones though, pick the low hanging fruit first, focus on those, then once you get some wins, you'll learn a lot in the process and then you can take that success and apply it somewhere else in even bigger investments, so to speak with whatever you're building, Automations, augmentations, doesn't matter. But again the key is to actually just build the thing. You can just build things now, the time to value is decreasing dramatically with things like cursor, vibe coding, et cetera. And you just do it.

Daniel Mishler:

Adopt the Nike motto here and just do it because if you don't, you're going to be behind.

Demetrios:

I love that philosophy and I want to talk to you a little bit about some of your learnings as you've gone and Vibe coded different investments. If we're staying on this VC theme and especially because I know that if you can get a very detailed PRD for what you're trying to create, you're two steps ahead in a way. And I also heard a friend of mine talking about how he likes to have a conversation and have the LLM ask him what he's trying to make and almost like suck information out of him and then it will go and it will help him put together the PRD after it feels like it's asked sufficient amounts of questions. I wonder if you have any insights on how to create a better PRD so that when you do go to Vibe coding something you are set up for success.

Daniel Mishler:

Yes, I think the PRD is absolute key to anyone who's Vibe coding. And trying to move quickly. You need some guardrails, especially like I found with Claude 3.7. If you don't keep that thing tightly focused, it can go and just make tons of different changes to make the best SaaS company or product in the world that you're not actually trying to build a SaaS product. But that's what it thinks it knows. But anyways, with the PRD specifically, my typical flow is I'll start previously when I first started Vibe coding and it's ironic because that term is a more recent term, but I actually started doing that about a year ago when cursor was kind of first taking off and I'd been doing the traditional thing. I tried to learn programming, never had time.

Daniel Mishler:

Like we just had our 10th kid. I don't have time for it. 10 kids.

Demetrios:

1 0.

Daniel Mishler:

1 0. Double digits.

Demetrios:

Wow.

Daniel Mishler:

But I just I never had the time. I had the desire. Never had the time to sit down and do Code Academy and everything like that. and you put a plop, a blank text file for me. I couldn't. I couldn't type.

Daniel Mishler:

I couldn't program anything. Maybe I could get to hello World on Python. Not sure but with something like Cursor, I quickly discovered, someone like me, no experience in coding, no experience In software engineering. I was able to make things that worked right. and to me, I'm big believer in the growth mindset.

Daniel Mishler:

That made me just fundamentally curious. what can I do? I don't. I'm too dumb and too inexperienced to know when to say no. that's not gonna work. So I just try it, And that's led to me just learning organically quite a bit. And I know I've gotten off topic a bit about the PRDs, but I'll get there, I promise. So I started with Cursor a year ago, not knowing anything. And the first thing I tried to build was a VS code extension similar to the Vibe coder that I just made a couple weeks ago.

Daniel Mishler:

It was kind of the same idea. can we maybe.

Demetrios:

It's good to say, what was the Vibe coder that you created real fast.

Daniel Mishler:

So I tried to take that idea of what you're saying of having the conversation in advance of actually sitting down and. And starting your Vibe coding. So I integrated the Vibe mode, where you can talk to a voice agent and just talk about an idea like you would, if you're just talking to your friend about it, generate a PRD automatically and then kick over. And then implemented a dictation with custom prompting within VS code as a VS code extension so that you could quickly just dictate using deepgrams, fastest, most accurate transcription. Of course, I'm heavily biased, obviously, but then, you could plop that right into Cursor Composer or Windsurf the Cascade and just start Vibe coding off the PRD you generated. So I had an early. The very first thing I started to build was a very early concept of that.

Daniel Mishler:

can I use my voice to command VS code? And so I was able to generate code with my voice just telling it to do things. I could do various VS code commands, unwrap or open a new File and call it this switch this file all with my voice. And it was really cool just to be able to see, again. I could actually make that. And I never dreamed of being able to do that because I just knew I was this is something I never prioritized early in my life.

Daniel Mishler:

I could prioritize it, but just maybe when I'm older and have more time, the kids are older, I'll have time again. But now, because of Cursor and Windsurf and these AI models, that's not the case anymore. Anyone can build anything. And so fast forward, I realized quickly again the importance of the prd. And big shout out to some Twitter personalities or ex personalities like riley Brown and McKay Wrigley. They were huge in just putting out resources. So that helped me kind of quickly as they were learning cursor. it helped to see other people validate kind of like what my experience was, where if you don't generate kind of a plan, you can quickly get off the rails and increase your chance of hallucinations.

Daniel Mishler:

But if you do this PRD ahead of time, then you have a much better chance of success. And so then my quickly developed, kind of my standard flow is. Well, before Vibe coder, I would just open composer or now it's agent in Cursor and just have that conversation, typing it out. And I would say, very importantly, don't code yet. Let's just talk about this. And then kind of bounce my idea. And then the nice thing is, I'd have this concept, but I wouldn't really know, as not a professional software engineer at all. I have this concept and this idea, but I don't know the exact technical terms to describe it.

Daniel Mishler:

But the nice thing is these models are smart enough where if you can point it in the right direction, it can guide you. And so that just kind of kickstarted my learning too, where now I learned again, I can't just sit down and type and write code, but I can at least, look through a code base and understand it. I've been told at a junior engineer level, despite no formal training.

Demetrios:

Wow.

Daniel Mishler:

And that's just because I'm doing it. I'm just learning, debugging things. Cursor, it makes lots of mistakes. but that's not the point. the point is if you approach it and you have good critical thinking and good ideas, you can. And you're curious, then you can go and actually go into the code base and figure things out and guide it a little bit better. Even if you don't know the exact syntax of what to type, you can at least guide the model to fix it for you. And so because of that and because of this new plan of developing PRDs, I've had a lot of success in building things and I've actually built some customer facing demos that have led to one deals for us.

Demetrios:

No way.

Daniel Mishler:

I don't know if I can talk specifics about them, But I never thought I'd ever do that with my background. But Cursor and Windsurf, this is no exaggeration, literally changed the course of my life and my career because of this capability it gave me, and you just have to be curious enough to do it. but anyways, that led to me then having this idea of Vibe coder that I love catching onto that term, like it's cool, it makes sense. I've been doing it before. Before it was cool.

Daniel Mishler:

One could say before the term was coined at least. And so I decided to take that agent, I hate typing. Typing slow, especially if you're just trying to have a conversation. So let's take that. At least put that in an agent within your VS code environment. And so now what it'll do is this extension. if you switch to Vibe mode, you can talk to the agent, tell it your idea, it'll critique it. You can even it's open source, so you can edit the prompt however you want, so go check out the GitHub repo, fool around with it, fork it, whatever you want to do.

Daniel Mishler:

But anyways, once you have that conversation, all you got to do once you're satisfied is at the end, just say, okay, based on our conversation, generate a prd. And then it actually will kick it out to a separate model that's a little more, a little smarter, O3. And again you can customize it so you can use O3 mini, you can use 01, whatever you want to do, 3.7, doesn't matter. And it'll generate the PRD for you automatically. And then from there you can switch over to the dictation and say, all right, here's the prd, get going and use your voice and dictate from there. So the thing with Vibe coder is this is just a preview though, Like this is how I think the world is Going to be where? Voice computing, what's the voice os? We're shoving voice into existing interfaces that are meant around screens, mouse and keyboard. But I think all that's going to radically change.

Daniel Mishler:

Why do we need to be tied to a big display, with things like AR glasses coming out or smartwatches, things like that. do you need a big display in front of you, or can you talk to an agent that's running your computer remotely and you're just seeing, a little bit of crucial information displayed on your watch or on your glasses and you're just conversing from there. And that's true vibe coding, It's now you're out at the beach, you don't even have your computer. You just have a smartwatch or AR glasses and a smartphone or something, and you're completely untied from that. And it's beyond just dictation, So you see a lot of apps that are putting in dictation, which I think is a great step. But this is like the next step beyond just adding dictation to an existing interface. And no one, no one's built that yet. I have ideas around it.

Daniel Mishler:

I'm going to fool around with it. But I think, to get back in this mentality of, things are moving so fast, the opportunities are massive. There's just so much green field out there that you just got to build things. Just build it, get it out there, ship it, get some feedback, drop the things that aren't successful, iterate from there.

Demetrios:

What have been some gotchas as you were building?

Daniel Mishler:

I think the biggest gotcha in cursor, at least for me personally, is the error loop, where you're pushing your context window and you run into an error and you're fix this. I gave you everything you needed. Fix this. And that's all you tell it to do. And then it'll fix it and say, everything's working great now. And then you test it and now some new error pops up that generated this error. Fix this.

Daniel Mishler:

And once you do that three or four times, you realize you'll get back to the original error, and so you can get into this endless loop if you're not, at a certain point, if you're designing more complex applications, a lot of people criticize cursor and say, you know, it's not going to be suitable for large actual applications and code bases and I disagree strongly. I think it will be. I've used it as such and I've made some complex things beyond just, a to do app or video games or a spinning hexagon with a ball in it. I think you just have to be aware of it, And so you have to understand that true, Vibe coding, only by Vibe is only going to get you so far. Eventually.

Daniel Mishler:

You do, need to be able to get into the code base. But that doesn't mean you need to know exactly how to code. You just need to know how your code works. You need to be able to think critically and understand the flow. And that might be a controversial statement, but that's what's worked for me. And again, I got my first ever PR request. PR actually in a production code base using Cursor. I did something to help the engineering team.

Daniel Mishler:

It was very isolated, it was an experiment on our end and it certainly could have gone better and I learned a lot from it. But like that was vite coding something into a production database. In production code base. Apologize if I'm using the wrong terms here. Again, still learning. But it is possible. It is possible and it's only getting better.

Daniel Mishler:

I think if you are listening to this podcast and you haven't used Cursor and you haven't used Windsurf, use them, try them, Use an open mind. Understand it's not gonna be perfect, don't expect perfection. But understand if you understand the limitations of the tool, you can use it very effectively. And there are lots of stories out there of people doing it on X, on LinkedIn and such. I think if you're not using that, you're going to be behind the eight ball. But if you do use it, you can speed up the tedious co development and truly focus on the thorny problems that certainly people like me would never be able to solve. That becomes the new moat for companies and software engineers.