Amogh Chaturvedi

Before building anything, we begged for calls. A hundred product people said yes.

The way we validated was just beg for calls. Just like, please talk to us. Please talk to us. I think even before we decided to build even a fraction of the product, we talked to over a hundred people easily who were not just like random people.

A whiteboard and tally marks decided what we built.

After every call, I used to go on the whiteboard and just write like four or five bullet points of what we learned. Four or five bullet points and then tally them based on like what other people were saying. So like if I sat on five calls and all five of them said the same thing, that thing was just like tallied up.

Don't start building until someone has signed a contract or paid you.

you should not build until you've talked to a bunch of people until you have like already like validation through like contracts or money that people want to buy this thing. And yeah, I think that's like the biggest thing.

If the problem is on fire, customers will forgive a broken product.

Like, even if your product sucks, one thing we noticed is like, if you're solving like a hair on fire problem, people are willing to forget about all the bugs, like your onboarding, like stuff just fails, whatever.

Everyone knew someone would build this. Vision models just weren't good enough until now.

It's just a matter of like Gemini wasn't good enough. Vision models just weren't there yet. I think the why now was just like really compelling is like models just got good enough to do this. Let's try it.

We pitched YC an accounting startup. Nobody on the team cared about accounting.

So we applied to YC with like an AI accounting thing. And like, what do I know about accounting? It just seemed like something like, oh, it's like the old space, like QuickBooks sucks. So we were just like, my co-founder was actually like a cracked accountant.

Two YC rejections taught him the same thing: stop hunting for ideas and notice your own pain.

But like, we were just in like the mindset of just like trying to find an idea instead of an idea organically coming to us. While we were building these things up, we installed like post hoc, all these session replay softwares.

Some customers send a million sessions a day. And the model hallucinates.

To like handle like millions of sessions coming in per like some of these customers have like a million sessions a day. Especially these like e-commerce sites that have a lot of traffic. That's been really hard. Gemini hallucinates quite a bit.

Someone used to sit and watch customer videos. Now an AI does it.

Normally people watch your session replays, right? We just want to get rid of the human part and just have an AI do it.

The YC interview test: can you say what you're building in two sentences?

The questions they ask are actually like looking back, like very easy. Like what are you building? Like if you can't answer that question in two sentences, like, like it's, it's going to be a problem.

The best thing about YC isn't the advice. It's not being alone.

I think building a startup is like quite lonely. It's just me and my co-founders most of the days. I think YC, the greatest asset is like the community, the network.

At Stanford we talked about launches and raises. Customers only care if you solve a real problem.

Because when I was at like Stanford, when you look at like startups and you talk about them, a lot of times, like what we talk about is like launch, how much you're raising, like branding, all like the sexy stuff. But in reality, like no one really cares. Like if you're not solving a real problem or whatever, like it doesn't matter how cool your website looks or like whatever, like it doesn't matter.

Most startups don't die from a bad idea. They die from co-founder fights.

I think the co-founder dynamics are like the most important thing. Like I think you'll see like some statistics. Like most startups fail not because like the idea sucks or the product sucks or whatever. It's just co-founder disputes. And that's the main reason.

Our AI was useless on a one-page website. It was only as good as its data.

I feel like at that point there isn't that much information. Like AI is only as good as the information you feed it. They had like a bunch of customers, like people like viewing their page. All we could do was just like figure out to the depth they scrolled and whether they clicked the button that led to the Stripe link