Hrishi Amravatkar, Mighto AI — Voice Is the New Interface for AI
Watch on YouTube, or listen on Spotify.
Hrishi Amravatkar has spent 20 years in the Bay Area, at eBay, Cisco, Nokia and Amazon, where he headed the conversational AI team on Alexa. The whole time, he kept building on the side: a used-book marketplace at USC, one of the first dating apps on the App Store (400,000 users), an early mobile telemedicine app. Now he’s the solo founder of Mighto AI, whose voice agents call contractors for property managers, gather quotes, negotiate and book site visits. Mash and Hrishi get into how he picks one idea out of a notebook full of them, where the moat is when anyone can build fast with AI, the agent that reviews its own calls overnight, and why he’s comfortable building alone.
This conversation has been edited and condensed.
“Everything around us is invented by someone”
Before we get into Mighto, tell us about your background — where you came from, and how you ended up where you are now.
Before Mighto I was at Amazon, on the Alexa team, heading the conversational AI team, and before that I was in Prime Video ads. I’ve been in the Bay Area for 20 years. I did my master’s in computer science at USC in LA, then came to the Bay Area and worked for companies like eBay, Cisco, Nokia and then Amazon.
Throughout that journey I was always an entrepreneurial kind of person who wanted to build things. At USC I built a platform for buying and selling old books; back in 2007 there were 33 universities on it. During my eBay time, in 2009, when the iPhone had recently launched, I wrote one of the first dating apps on the App Store — location-based social networking — with 400,000 users. At one point I was actually competing with Sam Altman’s company, Loopt, and we were sharing ideas about what we could do in that space.
Then I built one of the first mobile telemedicine apps. Everything at the time was very rudimentary and web-based. I wanted to make it much more handy, and onboard doctors onto the platform. That was a little too early, though.
It sounds like your brain fires with ideas all the time. Have you always been this way — seeing opportunities, or problems, everywhere?
This came from my childhood. As a kid I used to read books about inventors and scientists. How did Edison invent the bulb? How was the first car made? Everything around us is invented by someone, and it takes significant effort, critical thinking and inquisitiveness to see what can change the existing landscape. So I always kept thinking there’s always going to be improvement on the current world. Somebody built a bicycle, and then somebody built a car. Even in my undergrad in India, I built one of the first movie booking sites, around 2001. I just had to go for my master’s.
When we had coffee years ago, you told me you have a book full of ideas. If you’re inundated with ideas all the time, how do you choose the one to put your effort into?
Even for Mighto, it took me a year and a half to narrow down on an idea. There are multiple things. First, founder–market fit: do you have the background, experience and exposure for that idea? Second, do you have a personal pain point?
The telemedicine app came from when my wife was pregnant and we were having issues reaching a doctor. In India it was very easy to have a family doctor and just call that doctor at night or during the day, and most of those calls are informational. So I was like, why is it so hard for me to reach a doctor in the US?
Mighto is similar. I was managing a 100-unit property myself, and I kept facing vendor management: fetching a security agency, then an agency for gardening, or for roofing and construction. I felt this tedious work needs to be innovated with AI voice agents, and I had that experience from Alexa. So I have the pain point, and I have the expertise. Now it boils down to execution.
A lot of founders I work with say, “I have this pain, I think I can fix it,” and start building. But there are startups everywhere. What happens between identifying the pain and deciding you have to build it?
With vibe coding and AI, it’s way easier to build now. But that means you can end up in a trap — a feature fallacy, where you’re just building features and you’re not aware of what the customer needs.
So I was using different AI models to evaluate my narrowed-down ideas, with a SWOT analysis on each. I was also going through the startups that top incubators like YC are admitting in their current batches. You need to know the competitive landscape, otherwise you’re going to miss out; things are innovating significantly fast. I used to ask Claude, ChatGPT and Perplexity to rank my ideas against the last three YC batches: where might this idea fit?
That narrowed it down to three or four ideas. Then I added more factors: founder–market fit, go-to-market, execution, and the moat, or unique edge, of the idea. Right now, copying an idea from a technical standpoint is very easy. So you have to have some kind of network effect, some kind of unique edge.
Where the moat is now
In the old days you had to actually invent something unique you could patent and use to protect your technology. Where does that fit now, versus getting to market fast and building that network effect?
Companies that are purely developer tools are going to have a tough time, because companies can build developer tools themselves with a few prompts. That’s my personal opinion — take it with a grain of salt.
Ideas that have ground impact are different. In property management, a human interaction happens with the AI agent. There are two actual customers, and AI is augmenting that. That can create a brand and a network effect. And the more customers are onboarded to the platform, the smarter the AI gets.
Here’s a moat, for example. When calls are made from my platform to a vendor, I have a job that runs every night, or when every call finishes. Claude Code is running on my Docker container in AWS. It analyzes that call: what were the missing points, how can that call be improved? And it creates a PR for me to review the next day. So it’s enhancing itself as people use the platform more and more. Eventually I might have my own open-weight model, which gives me a much more unique edge.
But to answer your question, I think the moat is going to be more about branding, network effect and how quickly you execute your startup. These are utilities for you. Claude or OpenAI is not going to go into the vendor management space.
“It analyzes that call: what were the missing points, how can that call be improved on? And it creates a PR for me to review the next day.”
— Hrishi Amravatkar · Inspiring Founders Podcast, S2, Ep. 05
And you can lose a whole bunch of time focusing on research and invention, when you could get in front of customers and learn their pains really fast. Later you can train your own models and get that edge — but that comes from working with customers.
Definitely. What I’ve realized about AI is how significantly it reduces the resources you need. You know that from building Radi8: a single person can do 10 people’s jobs, right from design to implementation to execution. I’ve built four apps — native iOS, Android and web — plus the whole AWS architecture and the CI/CD. I’m building industry-grade, which I couldn’t have imagined for my previous startups.
“Everything is verbal, nothing is written”
Without divulging your secrets: Mighto isn’t another CRM or lead-gen tool. You’re connecting humans to humans to get work done, with AI in the middle. What’s the elevator pitch?
At Alexa I was working on conversational AI, and I realized AI agents are much more autonomous now. Previously we had answering machines doing rudimentary work — “press one, press two.” Now voice agents can handle complex, autonomous tasks with humanlike proficiency, and the scale is inhuman: they can talk 24/7, throughout the globe.
That’s where I got into Mighto. I was in that property management scenario — I was chairman of the society — and I had to call four or five vendors to get the best quote for the solar we needed to set up for the building. First of all, where do I get the contacts? Many of them don’t even respond. Then follow-up calls, WhatsApp messages, getting their site visit. All of that was very tedious. Everything is verbal, nothing is written, no audit trails. Eventually I realized: why not have an AI agent handle this complex, fragmented work of negotiating and talking with vendors? That wasn’t possible before.
On the consumer side, for property managers and even homeowners, there are iOS and Android apps. On the vendor side, I have almost 200,000 vendors onboarded right now, through the Google Places API. Those vendors don’t need to do anything, and they’re getting more business, because my voice agents give discoverability to vendors who weren’t getting reached before.
Is this part of the agent-to-agent world, where assistants talk to other assistants?
The way it works, a property manager enters the requirement through the mobile app, which I made as simple as possible. You record by voice: “I have a leaking sink, I need a plumber quickly. I’d like someone to come between 5 and 7 in the evening, Saturday or Sunday. If possible, ask them for a free inspection.” It creates the entire job for you, including your availability. You choose five vendors or 10, and press submit.
It sends SMS messages, and it reaches out to those vendors through an AI voice agent. You’re sitting at home. The agent says, “Hi, I’m calling from Mighto AI, I’m an assistant for Mash. We have a sink that’s leaking. Can you help us with the best quote?” The vendor asks for the address, who the owner is, and it provides all that information.
Wait — that’s really important. Using voice to input data isn’t new; I can tell ChatGPT or Claude, “I need a plumber.” What you’re saying is that your system continues the voice all the way through to the vendor. They don’t get an email or a lead through some CRM. They get a phone call with an agent, and it goes back and forth.
Exactly. It’s streaming over WebSockets through the LLM, all at runtime, so latency and inference are extremely critical. You cannot have an agent thinking for multiple seconds. So I’m using Groq Cloud, and there are others, like GPT-OSS 120B — the fastest models for use cases like voice agents. There are different models providing voice activity detection, and barge-in detection, for when someone speaks in the middle of your speaking.
On the stream, you’re intercepting keywords. Say the vendor says, “I want to do a site visit.” “Site visit” is a keyword, caught by a regular expression; it doesn’t even need to go to the LLM yet. It immediately checks the user’s availability — Saturday or Sunday, 5 to 7 — and asks the vendor what time they want. The vendor says 6:00 Saturday, and a push notification goes to the user to confirm. If you confirm, it’s confirmed on the same call; if the call has ended, the vendor gets an SMS.
If the vendor quotes a price — say $5,000 — and I had asked for a 10% discount, it’s going to negotiate: “Can you do 4,600? 4,800?” If the vendor says no, it asks whether they can do a free site inspection, because the user asked for one. Then the user gets a push notification: these are the vendors you’ve got quotes from, these are providing free site inspections. All you have to do is choose which vendor you want to go with — or call them directly.
This actually solves two problems. The property manager needs a bunch of quotes; the tradesperson wants more work and more leads. It’s a double-sided marketplace done through voice — and people don’t have to join, because you’ve already put them all in there.
Yeah. All it takes is adding a new subcategory. You and I had that discussion about your tree service scenario — tree cutting is a subcategory where people might want vendors. I just added tree services for San Francisco, LA and New York. It’s very easy to integrate those vendors into the platform.
“Machines are building the building for you”
It sounds like you’re building this by yourself. It used to be you needed teams and teams, investors, backing money. How are you working today?
I’m building it as if it’s going to be an actual profitable business. Investors come or not, it has to be a viable business. And you’re right, I’m a solo founder. That wasn’t possible for any of my previous companies, at this speed. My background in engineering and architecture has helped me build this application very fast and to industry standard.
I used Claude Design for designing the app to begin with. Then I created product requirement documents. It’s like building a building now, where machines are building the building for you. But you have to do the architecture and the design, and make the foundation strong enough that they can follow on top of it. So initially I was way more micromanaging, so that Claude understood what I needed — through MD files, product requirements, granular details, even correcting code. Eventually it matured enough to be almost autonomous.
So far, solo. The plan is to eventually hire for go-to-market and execution, because I’m busy now. But I think it’s much easier to be solo.
If you don’t raise money from a VC, you don’t actually need to become a billion-dollar company. There’s probably a much more achievable number that would make you immensely happy — a hundred good customers, or a thousand good properties, using the service. Is that your thinking?
Yeah. Personally I think scaling can be done easily with minimum resources right now. Second — and this is more personal advice for everyone — being in a job is way more complicated, and much riskier. When I’m building, I’m not employing that many people. The value of employees has significantly reduced. So it’s good to be an entrepreneur at this stage, and use those tools to amplify your knowledge and your entrepreneurial side as quickly as possible.
There’s another model emerging, where solo founders build apps that get acquired quickly. If you show up with a thousand property management companies and real monthly recurring revenue, it’s cheaper for an enterprise to buy you than to build the whole thing with multiple developers.
It is much harder for a bigger company to build this than a smaller company like mine. If I have 10 engineers, they’re all going to think in a different fashion, and I have to keep them in sync. I’ve seen a simple change have to go through multiple bureaucracies — when you could just upgrade the model once your evals are successful. The bigger the company, the slower the innovation pace. So, as you said, there is a possibility of that acquisition scenario.
“Customer first”
Let’s talk about your vision. The technology piece, you’ve nailed. What are your ideas for go-to-market — how do you get these property managers?
For me, the product has to be built with the customer in mind — customer first — and my customers are on both sides: property managers and vendors. I’ve built the product multilingual. In the US it can speak Spanish or English on the fly, so if your vendors speak Spanish, it can communicate with them. It’s also getting launched in India, where it can speak 11 or 12 languages and communicate with vendors locally. On the ground, it should feel valuable, comforting and understandable for the vendor side as well.
Initially my plan is to target B2B, because then I can bring in many more jobs, or RFQs. A property manager has multiple properties, and all of them require some kind of job almost every day. In B2C, the cost of acquiring customers is going to be very high from day one. So it’s better to do B2B, and then move to B2C.
You’re also solving a problem for contractors — getting them leads. How do you think about that side?
Right now property managers are using the platform to get their stuff done as fast and as smoothly as possible. They’re saving almost 10x of their time — they can reach 10 or 20 vendors without reaching out themselves — and they can save 20 to 30% through the platform, because they’re reaching out to multiple vendors.
Phase two of the product is on the vendor side. Right now vendors aren’t onboarded on the platform per se. They get a voice call, an SMS and a unique web URL where they can update the quote. But I’d like vendors to actually be onboarded too, with a verified tick mark.
You might have seen companies going into vendor-side AI answering machines, and they’ve raised a significant amount of money. Vendors upload their catalog of services and customize the voice agent, so when a customer calls a Mac repair shop, it asks, “Is it an iPad?”, says what it will cost, and coordinates. Those companies charge a pretty hefty amount. My goal is to charge almost half, with additional benefits like invoice management — and my answering machine is going to have all the catalogs and the pricing. Eventually I’d like my AI voice agent on the property manager side to talk to the AI agent on the vendor side, and the two of them negotiate a price for the business.
“Learning to take rejection”
You came all the way from being a kid curious about inventing things, through companies like Amazon and eBay. If you could share one thing with a new founder, given what you’ve learned about building with AI, what would it be?
There are a couple of things: one is on building the product, another is on execution.
Building the product has definitely become much easier, but I’d suggest founders stay way more grounded in the technology. Don’t have everything done autonomously through AI agents, especially in the initial days, because it can overengineer things. You end up with a huge blob of code across thousands of files, and an agent makes changes in one part without realizing something else will break. Eventually you have to get into the weeds anyway, so it’s better to do it right from the beginning, and then let the AI become autonomous.
I’d also urge them to use the new techniques — creating skills, creating routines — not only for building the product but for the other tasks founders need to do. I have a routine that runs every morning on my interviews: Loom has the recordings, and it transcribes them, updates my Notion doc and gives me a rating. The whole job of a recruiter is now done through a Claude routine or skill for me.
On execution, you need to be a good storyteller. You need a good narrative for your product: why you’re building it, how you’re building it, and how it solves a real problem. And as engineers — I’m a culprit too — we focus a lot on building products and innovation, and we’re sometimes way weaker at talking to customers. So eat your own dog food. And as a founder, reach out to customers yourself rather than quickly hiring a sales guy. I’m not hiring sales until I’ve talked to a few customers myself and gotten their feedback.
And always be open to rejection. You’ll hear a lot of nos. Many people will say, “What is this? It’s just an answering machine,” and feel it might not work out. Even investors will say, “You’re too old to do a startup — there are youngsters out of college doing startups.” I’m not going to take it personally. I’ll keep on improving on top of it. So: learning to take rejection.
A lot of those skills aren’t engineering skills; they’re product management skills. You have all the pieces. What would you say to someone who doesn’t? A business student used to say, “I need a technical co-founder,” and an engineer, “I need a business co-founder.” Do you still tell them to find a partner?
There’s still a use case for co-founders with complementary skill sets. But if a CEO is a little more tech-savvy now, they don’t need a CTO as a co-founder as early as before. Less equity gets given out, because AI is doing the product building, and you don’t need a VP of engineering from day one. Design and sales, I definitely feel you need. If you have someone with a sales background coming in as a partner or co-founder from day one, that’s good.
And the initial calls are all customer discovery anyway, so you and your partners will spend an inordinate amount of time interviewing people and understanding their problems.
The main thing I feel is this. There were times in my own career when we had two technical co-founders — me and my co-founder. It was complicated to build that product, and we needed to keep motivating each other. That’s good if everything works well. But the drawback, which everybody knows, is co-founder differences. Instead of focusing on execution, the company focuses on one co-founder wanting to leave for a high-paying job, or personal life issues. That baggage comes with a co-founder who has the same skills as you. So in this new age of AI, if you hire, hire complementary skill sets, way more focused. And it’s easy to be a solo founder as well.
“Why do I need to hire someone?”
Congratulations, and thank you for sharing so transparently.
Thank you so much. And thank you for helping with the marketing and branding through Radi8. It’s been a great experience to use it as a routine, AI-based agent that helps me on the side of the business where I’m weak.
Since you brought it up: what problem did you have when I introduced it to you, and is it solving it?
It was serendipity. I was actually facing the problem: how do I start building awareness for my product before it launches? And when it launches, I need some kind of social media outreach that runs consistently. I was thinking, why do I need to hire someone to do that — a full-time salesperson to do this marketing? That’s when I saw Radi8. I knew you’d been building it, so I started using it and deep-diving into it.
Initially I wondered whether it would be able to extract the right content and create the right kind of posts for me to publish. I was really surprised by the quality of the content it’s creating. Quality is extremely important, and I’ve hardly had to update or change anything in the posts Radi8 creates. There’s engagement on those posts as well. Right now it’s posting on LinkedIn, and I want to connect the other feeds too: Twitter, Facebook, Instagram. I’m looking forward to using it more.
More conversations with founders on the Inspiring Founders Podcast.