TL;DR

• AI removed our engineering bottleneck, which exposed new ones in design, product thinking, and feedback loops.
• We learned to ship earlier, use constraints to stay focused, and rely on retention and real behavior as the clearest signs of product market fit.
• That clarity pushed us to build Radiant as a standalone product, even though Supernormal had already reached 700k companies.

---

I’ve spent the past few years building AI products with my co-founder, Fabian Perez. Together we run two products, Supernormal and Radiant. Both came from the same goal, helping people work faster, with more clarity, and with less of the drag that usually follows every meeting.

We launched Supernormal in 2020 and it grew quickly. Faster than anything I had built before. 700k companies used it, and many still use it today. Supernormal started as a simple idea. Make meeting notes effortless. It automatically captured conversations, summarized them, and helped teams stay aligned across projects and decisions.

As adoption grew, we started seeing a different workflow challenge emerge, one that didn’t sit neatly inside a meeting notes product. Supernormal continued to thrive, but extending it to solve this new problem would have stretched it beyond its purpose. To build what we had in mind, we needed a new product, a new identity, and a clean slate.

Here is how we made that decision, what AI changed for us as builders, and what I learned that might be useful if you are building your own product or company, and wondering how and when to pivot.

AI changed engineering speed, and suddenly everything else was too slow

A few years ago, engineering used to be the slowest part of the process. Before Supernormal and Radiant, I led product teams at Facebook and Klarna, and it was normal for engineering to set the pace.

Then AI arrived, and overnight, our engineering velocity jumped by roughly 3–8x.

That sounds great, and in many ways it was, but it created a new bottleneck. Design and research did not scale at anything close to the same speed. We could build a working AI feature in a day, but designing it well, testing it with users, and getting feedback still took much longer.

Mockups stopped being enough. They were too static for AI behavior. To understand how an AI feature actually worked, we needed to test real, functional versions with real data. One thing I learned quickly is that AI features cannot be evaluated in isolation. You only get useful signals when the full product is running end to end, with the real data flows and real context users experience. Anything less gave us misleading feedback.

That forced us to rethink our entire loop.

We started to:

• Ship functional prototypes instead of polishing Figma files forever
• Tighten feedback cycles so we could see how real users behaved, not just how they reacted in a demo
• Add deliberate pause periods so the team didn’t burn out by operating at maximum speed all the time

AI removed one constraint and exposed several others.

The mistake I see founders make with product market fit

As Supernormal grew, one lesson became clearer. A lot of founders scale the wrong thing at the wrong time. Some chase growth before the product resonates. Others refine endlessly without knowing who they are building for. I have done versions of both myself.

The most reliable signal for us was retention. For impactful workflow tools, 50–80 percent retention usually means you are solving something meaningful. Then qualitative feedback tells you what is actually working.

When I say qualitative feedback, I mean:

• What people say they would miss if you took the product away
• Where in their workflow they naturally slot in your product
• Whether they are trying to bend your product into something else

Behavior always tells the truth faster than opinions.

Why Radiant needed its own space

Radiant began as a set of ideas inside Supernormal, but the more we explored them, the clearer it became that we were not building a better notetaker. We were building something in a different category entirely.

Supernormal is a meeting notes tool for teams. Radiant is an AI assistant and workspace for your Mac that sits much earlier in the workflow and goes far beyond notetaking. It captures your meetings on device, generates summaries and suggested next steps, and helps you actually move that work forward by drafting emails, updates, and documents. It is designed to help individuals act on their meetings instantly, not just remember what happened.

Keeping everything under one brand would have created confusion and slowed both products down. Spinning Radiant out as its own product and identity gave us the freedom to:

• Build a clearer story
• Focus on a sharper ICP
• Explore product ideas that didn’t fit Supernormal’s constraints

Supernormal stayed focused and continued to grow. Radiant got the room it needed to become something new.

Early traction came from user behavior, not predictions

When we started testing Radiant, the validation was immediate. People didn’t just say they liked it. They kept using it. They opened it after meetings without being reminded. They pulled more of their workflow into it.

That behavior gave us confidence faster than any interview or survey could.

We shipped imperfect versions early. Live meeting note generation was not flawless. The interface needed refinement. Some flows were still rough. But the core value was real, and shipping early helped us tighten the loop between concept and improvement.

We invested in infrastructure that let us move fast for the long term:

• Better component libraries
• Faster prototyping
• Design systems that made iteration cheap instead of painful

Constraints made the product better

One of the most defining choices we made with Radiant was deciding what not to do.

Radiant does not use a bot that joins your meeting as a participant. People dislike that experience, and it is expensive to operate at scale. Instead, Radiant captures audio directly on your Mac and processes it locally to generate transcripts and summaries. That constraint shaped everything, from our data models to the way the product feels to use.

Constraints didn’t limit us. They created clarity.

What building two AI products taught me

If you are building an AI product today, you have very little time to win someone over. Users give you about thirty seconds to understand what your product does and feel some kind of value. They have endless choices, and the moment they hit friction, they churn.

Here are the lessons that helped us build both Supernormal and Radiant in that environment:

• Build using real data, not static ideas
• Keep your ICP tight
• Ship early and iterate fast while you still have momentum
• Watch what users do, not just what they say
• Use constraints to create focus
• Protect your team so they can maintain a sustainable pace
• Let retention and behavior guide your next moves

Supernormal and Radiant are different products, but they share the same foundation. They were shaped by watching how work actually happens, staying close to users, and adapting quickly as the AI landscape shifted under our feet.

If you are building in this space, or considering spinning out a second product, I hope this gives you some useful signals and a few shortcuts.

Follow along

If you want to keep up with what I’m building, I’m mostly on LinkedIn. You can also follow Radiant and Supernormal on all the usual social channels, and download or signup on their websites. And if you ever want to swap notes about building products, please get in touch.