Saved copy: How AI startups can deal with massive data, with Diego from Krea.ai

[andyvan-ph]

`

The irony of going viral is that everybody wants it, but nobody’s ready for it.

There is often such a sudden increase in data usage that servers strain, performance degrades and costs can shoot up for founding teams.

The team at Krea surfed that wave — when their app went viral in Southeast Asia, hitting 200,000 sign-ups in a single day. Having already set up PostHog to track events for Krea, they were ready in advance for such a surge.

Diego Rodriguez, cofounder at Krea, senses that startups usually don't need to care about big amounts of data until it’s too late but that way of working may change soon if it didn't do it already.

His advice? If you’re building an AI-first app, you’re going to
need to be ready for massive surge of data, independently of whether virality is in the equation or not. Here’s how founders can prepare with a stronger foundation.

1. AI-first apps generate more data than you can dream

Generative AI presents an entirely different paradigm for managing data such as events. “When you track user events on a finance website,” Diego says. “People don't go to a website and spend money a thousand times on a visit. That just doesn't happen. But people do come to Krea and they may make 1,000 images in a single session.”

Diego explains: “Let’s say your product is Canva. How much data is passing through your product? A user creates one artifact: let’s say, a poster. There it is. They’ll create one poster. You don't make many posters per second, or many posters within one minute. You make A poster. You work on it, and then you finish and then you save it. But in the world of AI?

For me, the equivalent of an artifact that the user made it is not a poster — it’s a frame that you make on the Krea Realtime tool. But the thing is, within one minute you can make a lot of images, right? Yet, each frame is an artifact of a kind — one moment or “event” in terms of product analytics. The person is making images, but suddenly you are not processing few of them in a few seconds; instead, you are processing many frames per second during, say, 20 minutes. We are not talking about percentage differences anymore—it's in the orders of magnitude.”

2. Analytics needs your full attention

Unlike conventional apps, the data generated by apps built on generative AI will compound. Whether or not you grow users significantly, your events are always going to keep growing.

As a founder, it’s true that your primary concern should be the product you’re making. But there are three other systems Diego advises to pay close attention to: your billing logic, infrastructure observability, and analytics. “You should always think of the analytics repercussions of each one of your product launches. At least the basics.” he insists.

“Besides,” Diego brings up, “independently of what your app does, analytics is the foundation of understanding. It's one of the ways to understand your customer, and as a business owner, that is a core competency. That is industry invariant. AI or not, understanding your customer is a must."

"Contemporary businesses operate at a global and digitally interconnected scale today; this is not something exclusive to Google or "Big Tech" anymore. Two ears are fine when you need to listen to one person who speaks your language, but what do you do when you have thousands who speak different languages and from different cultures and backgrounds? For me, that's talking and listening to customers over social media in the foreground, while leaving PostHog deal with analytics nightmares running in the background."

3. LLM analytics are the new frontier

Diego had one last piece of wisdom to impart: pay attention to what’s needed in the realm of LLM analytics. How are you understanding your customer’s interactions with your AI models? Are you able to keep track of the LLMs themselves? What do they perceive? How are they interacting with your business and brand? Where once customer insights were confined to the realm of “how people found your app,” we’ve now got a new frontier of “what happened in the LLM that influenced a customer?”
In other words, early-stage products used to spread via word of mouth; today, we will start seeing them spread through word of... next token prediction?

(PostHog has just launched a new beta product in this category, introducing LLM observability to let you check in what took place in the deep underground pipes of your in-house agents and AI features.)`