Rohan Joshi grew up in a Mumbai home where nobody said the word “entrepreneur,” but everybody lived it. His father builds pharmaceutical machinery, his mother runs a salon chain and her own cosmetics brand. So when his parents told him to “go get a job for experience,” he already knew what the experience was really for: a checkpoint on the way to building his own company.
In this episode of Unstarted x Intelligent Indians, shot in the San Francisco Bay Area, Rohan sits down with Avneesh to talk about the story behind Stellar Labs (better known as KittenML), an AI research lab building open-source voice models so efficient they run on a CPU.
They get into getting deep in AI before it was cool, computer vision at CMU, and three years at Meta building photorealistic avatars. Why he and his co-founder didn’t wait for permission to start building. The honest experiment that broke their first idea: pricing a local ElevenLabs alternative at full price and watching every user churn.
The pivot to developers, open-sourcing a voice model, and watching it go so viral on Reddit they assumed their own analytics were broken. His framework for knowing when to build instead of doing customer research. The three inflection points coming for voice AI. And the mindset that ties it all together: the easiest person to fool is yourself, so fool yourself into believing you can do anything.
If you want to build in AI, this one is a playbook on conviction, shipping fast, open source, and why the five-year grind that used to separate the experts is now something AI hands you for free.
0:00 Intro: an AI lab that runs on a CPU
1:20 Growing up in an entrepreneurial Mumbai family
2:10 "Do a job for experience": the unspoken goal
3:15 Getting into AI before it was cool (CMU, 2019)
4:30 What actually changed: data, GPUs, and the transformer
7:00 From medical AI to Meta's photorealistic avatars
8:20 Starting Stellar Labs with a co-founder from Meta
9:30 Why they didn't wait for permission to build
10:40 Conviction vs delusion: should you do customer research?
12:00 Why the biggest companies started before they were businesses
13:30 The first idea: a local alternative to ElevenLabs
15:30 The experiment that broke it: everyone churned
16:40 Why open source opens a bigger market than closed
18:20 How do you know it's working? When you doubt your own analytics
20:40 The future of voice AI: three inflection points
24:30 Compounding: what looks like talent is years of work
26:40 Voice AI that runs on a $20 server
28:00 Should you move to the Bay Area to build?
29:50 Parting advice: fool yourself into believing you can