It started in 2015 with Unity.

I was building games — indie, scrappy, learning-as-I-went. Unity had just opened up in a way that made real-time 3D accessible to people who weren't at EA or Ubisoft. The tools were raw. The documentation was spotty. Half the forum answers were wrong.

But something was happening. A new layer of computing was opening up, and the people who understood the primitives — meshes, physics, shaders, scene graphs — were suddenly valuable in ways they hadn't been six months earlier.

I didn't know it then. But I was watching a platform shift from the inside.

The Second Time

2018. I moved into AR/VR.

HoloLens had shipped two years earlier to developers — raw, $3,000, no consumer release in sight. Magic Leap was still months away from its own launch, and would go on to disappoint nearly everyone who'd spent years anticipating it. And yet there was this genuine electricity in the air — the feeling that spatial computing was about to become the interface. We were building for it before most companies had a strategy for it.

The tools were, again, raw. The use cases were, again, unclear. The people who'd been doing this for two years were, again, suddenly very valuable.

I noticed something: the early stage of a platform shift feels identical every time. Breathless hype. Murky tooling. A handful of genuine believers and a much larger crowd of tourists. And a window — maybe 18 months — before the platform either matures into something real or collapses into a cautionary tale.

AR/VR matured slowly. Slower than the hype. But the people who'd built real things in that window carried something forward that couldn't be faked: intuition for how spatial interaction actually works. When you've made a hundred decisions about where to place a UI element in 3D space, you stop theorizing about it.

The Third Time

2020. I joined Avataar.

The brief: build an AI platform for 3D commerce. The underlying bet: agents — software that perceives, decides, and acts — were about to become real infrastructure. Not demos. Not toys. Real, deployable systems.

I spent five years there building exactly that. We built the Core Agent — an AI system that could understand a 3D product, reason about a customer's context, and orchestrate a personalized commerce experience end-to-end. Multi-agent systems, RAG pipelines, orchestration layers — before most product teams had decided whether to take LLMs seriously.

By the time the rest of the industry caught up to what agentic AI could do, we'd already made most of the wrong decisions, learned from them, and rebuilt the right way. That's the only path through an early platform. You can't shortcut it.

The Fourth Time

2025. I joined Neoflo.

Different domain — finance operations, CFO teams, accounts payable. Same underlying moment. AI was being bolted onto enterprise workflows by people who understood AI but hadn't spent much time with the workflows. The opportunity was obvious if you'd seen this movie before.

My first week, I sat with an accounts payable manager. She showed me her screen: 47 browser tabs, 3 spreadsheets, an ERP that looked like it was built when Blackberry was cool, and a folder on her desktop called INVOICES_FINAL_FINAL_v3.

She wasn't bad at her job. She was good at it. She'd built an elaborate system of workarounds to compensate for software that had never been designed around how she actually worked.

That's the real problem. Not "finance teams need AI." The problem is: the tools don't match the workflow. They never did. AI is just the first technology with a real shot at changing that.

The Pattern I Keep Noticing

Four platform shifts. Different domains. Same sequence, every time:

Phase 1: The primitives open up. Something that used to require deep specialization or institutional access becomes buildable by a small team. Unity democratized 3D. Cloud democratized infrastructure. Transformers democratized language models.

Phase 2: The tourists arrive. Everyone with a Twitter account has a take. Conference talks get booked nine months before anyone has shipped anything real. Breathless LinkedIn posts about the "revolution" start outnumbering actual products.

Phase 3: The tooling is terrible. The documentation is wrong. The APIs change without notice. The "right way" to do things doesn't exist yet because not enough people have tried enough things. Building during Phase 3 is frustrating and expensive. It's also when you build the intuition that can't be acquired any other way.

Phase 4: The workflow problem becomes obvious. This is the one most technical builders miss. The hard part isn't the technology. It's understanding the human workflow deeply enough to know where the technology actually fits. In every domain where the workflow is complex and deeply human — finance, healthcare, legal, support — the builders who won weren't the ones with the best tech. They were the ones who understood the workflow first.

Phase 5: The window closes. The platform stabilizes. Best practices emerge. The tourists have left. The people who built through Phases 3 and 4 now have a durable advantage — not because they hoarded knowledge, but because they have scar tissue.

What This Actually Means

I'm not writing this as a victory lap. I'm writing it because I keep having the same conversation.

Someone asks: "Is it too late to build on [current hot platform]?"

The honest answer is: it depends where you are in the cycle. But the more useful answer is: if you're asking whether it's too late, you're probably already thinking about it wrong. The question isn't timing. The question is whether you're willing to spend time in Phase 3.

Most people aren't. Phase 3 is painful. You have to make decisions without enough information. You ship things that don't work. You rebuild. You explain to stakeholders why the thing that looked easy on the demo is actually hard in production.

The people who sit in Phase 3 long enough come out the other side with something nobody can take from them: they know where the bodies are buried. They've seen the edge cases. They have intuitions that took a hundred decisions to build, and those intuitions compound.

That's the pattern. The technology changes. The cycle doesn't.

One More Thing

The irony of building through multiple platform shifts is that each one makes you less surprised by the hype of the next.

When everyone was losing their minds about the metaverse in 2021, I wasn't cynical. I was just... calibrated. I'd seen how long it takes for spatial interfaces to become usable. I knew what "early" actually looked like.

When people started declaring that AI agents would replace all software by 2024, I wasn't cynical about that either. I was building in it. But I knew we were in Phase 3. The hype was running three years ahead of the tooling.

We're still in Phase 3 on most of what people are calling "agentic AI." The primitives are real. The use cases are real. Some workflows — customer support, code generation, search — are maturing fast. But enterprise deployment, especially in domains like finance and operations where accuracy and auditability are non-negotiable, is still far behind the hype. The tooling is still catching up, the best practices are still being written, and the workflow problem is still wildly underestimated.

Which means there's still a window.

The question is whether you're willing to sit in the uncomfortable part.