AI as a Feature, Not the Product

Most AI products are just chat windows with a logo slapped on top.

I don't mean that as a hot take. Open Product Hunt on any given Tuesday and count how many launches are a text box, a send button, and a wrapper around the same API everyone else is using. Some of these are genuinely useful. But the wave of "AI-first" apps has created a weird blind spot in how we think about building software: that AI should be the thing you interact with, rather than something that quietly makes the thing you're already using better.

I've shipped two apps recently that use AI in fundamentally different ways. Neither of them is an AI product. Both of them are better because of AI. And the distinction matters more than most people realize.

What "AI as a feature" actually means

There's a spectrum. On one end, you have products where the AI is the entire interface. ChatGPT, Claude, Midjourney. You talk to the AI, and the AI talks back. The AI is the product.

On the other end, you have products where AI runs invisibly in the background. Spam filters. Auto-exposure on your phone camera. Recommendation algorithms. You never think about the AI because it never asks you to.

The most interesting space is somewhere in between. Products where the AI is present, available, and genuinely useful, but it's not the main event. It's a feature that amplifies what the product already does well. The user came for the product. The AI just made it faster, smarter, or more accessible.

That's the space I've been building in.

Molecular: an AI tutor that knows what you're looking at

Molecular is a 3D chemistry playground I built with Three.js, React Three Fiber, and Next.js. Students drag atoms together, snap bonds, run reactions, and explore a library of 30+ curated molecules in real-time 3D. It's a learning tool first.

The AI tutor is one feature among many. But it's the feature that makes everything click.

Asking more questions about the current scene on Molecular

Here's what makes it different from just bolting a chatbot onto a chemistry app: the tutor is scene-aware. Every time a student asks a question, the current state of the 3D scene gets serialized and sent along with the prompt through the Vercel AI Gateway. The tutor knows which molecule you're looking at, what mode you're in (Explore, Build, or Lab), and what just happened. Ran a combustion reaction and want to know why it produced water? The tutor doesn't give you a generic answer about combustion. It references the specific reactants you combined and the products that appeared in your scene.

Context is everything. A blank-slate chatbot would be dramatically less useful here. Students would have to describe what they're seeing, specify the molecule, explain the context. That's friction. And friction kills learning.

I also built education tiers into the system: Beginner, Standard, and Advanced. The tier shapes every response the tutor generates. A middle schooler asking about covalent bonds gets a different answer than an AP Chemistry student asking the same question. Same AI, same model, completely different output because the context is different.

The tutor can even highlight specific atoms and bonds in the 3D scene while it explains them. It doesn't just tell you where the oxygen is. It shows you.

And the suggested prompts update based on what you're doing. If you're in Explore mode looking at a benzene ring, the suggestions are about aromatic compounds and ring structures. If you just ran a neutralization reaction in the Lab, the suggestions shift to acids, bases, and pH. The AI meets you where you are.

None of this required inventing new AI capabilities. It required thinking carefully about context: what information does the AI need to be genuinely helpful, and how do you weave it into an experience that already works without it? Molecular is a fully functional chemistry tool with the tutor turned off. The AI just makes it better.

PitchTrack: speed to data through natural language

PitchTrack is a completely different kind of app. It renders MLB pitch trajectories in 3D using Statcast data. You can compare pitcher arsenals side by side, replay at-bats pitch by pitch, and explore the dataset with a deep set of filters: pitch type, batter handedness, count, zone, season, outcome, batted ball type. The explore page alone has probably 40+ filter options across six categories.

That filter system is powerful. It's also a lot of UI to navigate.

PitchTracker's AI quick analysis feature

Say you want to see all of Tarik Skubal's sliders that got a swinging strike against left-handed batters during the 2025 season. Without AI, you'd need to search for Skubal, select "Slider" from the pitch type filters, check "Swinging strike" under outcomes, toggle "L" for batter stance, pick the 2025 season, and hit search. That's touching four or five different controls across the page.

Going directly to filtered results using AI to consolidate natural language to a analytics result

With the natural language feature I'm building, you just type that sentence. The AI figures out what you mean, maps it to the right filter parameters, and constructs the visualization URL directly. No chat interface. No back-and-forth. You describe what you want to see, and PitchTrack takes you there.

This isn't about the AI controlling the visual layer or generating novel analysis. It's about speed of interaction with the data. The filters already exist. The visualizations already work. The AI just removes the friction of getting to the exact view you want. One sentence instead of five clicks.

The important thing: the AI doesn't replace the UI. It's a shortcut through it. Power users who love clicking through the filter panel can keep doing that. Casual fans who just want to see a specific pitch can skip straight to it. Both paths land in the same place.

The restraint problem

Here's the part nobody talks about: knowing when not to use AI.

I'm constantly tempted to add AI in more places. What if Molecular's AI could generate custom quizzes? What if PitchTrack's AI could write scouting reports? What if every page had a little sparkle icon that summoned a helpful assistant?

I've talked myself out of most of these ideas. Not because they're technically impossible, but because they'd add complexity without solving a real problem. The quiz feature sounds cool until you realize students already have teachers assigning work. The scouting report sounds impressive until you realize the people who want scouting reports want them from scouts, not from a language model hallucinating about pitch tunneling.

Restraint is a design skill. Every AI feature you add is a feature you have to maintain, explain, and defend when it gives a weird answer at 2am. The bar should be high.

My filter: does this AI feature solve a problem that the user actually has, right now, in this specific context? Or does it just feel like something an AI-powered app should have?

If it's the second one, I don't build it.

Start with the problem

The pattern I keep coming back to is boring in the best way. Start with the user's problem. Build the product that solves it. Then ask: is there a spot where AI can reduce friction, add context, or make this faster?

If the answer is yes, build the AI feature with as much context as you can give it. Scene state, user preferences, education level, current filters, recent actions. Generic AI is a dead end. Context-aware AI is where the real value lives.

If the answer is no, leave AI out. The product doesn't need it just because investors expect it or because the landing page looks better with "AI-powered" in the headline.

The apps that will win the next few years aren't going to be the ones with the most AI. They'll be the ones where you forget the AI is even there, because it's so tightly woven into the experience that it just feels like the product is smart. That's the bar. That's what I'm building toward.