Why ChatGPT Can't Tune Your Car (And What I Built Instead)

About a year and a half ago, I had an Audi RS5 sitting in my driveway and an itch I'd had my whole life.

I've always wanted to be into cars. Growing up, I never really had a car that warranted modifications. The RS5 was the first one I did. And once I decided I was going to mod it, I had the same problem every enthusiast has: there are a thousand parts, a thousand tunes, a thousand opinions on every forum, and you have no idea which combination is actually right for your car.

So I did what a builder does. I opened ChatGPT and asked.

ChatGPT was my main AI tool at the time. Claude hadn't become the default for me yet. And ChatGPT was helpful. It just wasn't right.

It would give me generic parts. It wasn't specific enough to my car, my goals, or the tune I was considering. Worst of all — and this is the thing that still drives me crazy — it would stack performance claims. If part A advertised +20 horsepower and part B advertised +15, it would tell me to expect +35. Which is not how this works. The gains are shared across components. You don't get to add them up. Any car guy knows that. ChatGPT didn't.

Meanwhile, every time I'd hit a wall, I'd text my buddy Cory.

Meet Cory

Cory is a car genius.

I mean that literally. If you put him in a garage with any car — any generation, any engine, any combination of aftermarket parts — he will know what's in it, what's wrong with it, and what it needs next. He's a mechanic, he builds race cars, and he's been wrenching since he was a kid. He's also one of those rare people who can explain what they know to someone who doesn't know it yet — which, for me, was everything.

Cory had always wanted to start his own automotive company. Performance service center, tuning shop, something in the aftermarket space. He had the idea. He had the knowledge. What he didn't have — and he'd tell you this himself — was the activation energy to start. Getting something off the ground is its own skill, separate from being great at the thing you'd build.

That's more or less my superpower. I have an idea, I want to start executing on it ten minutes later. Classic complementary pair.

On the RS5, Cory helped me do everything myself. Cold air intake. Downpipes with a cat-back exhaust. Resonator delete. New exhaust pipes. A Stage 1 tune. Every bolt, every decision, every part selection — he walked me through it, I did the work, and when something went sideways, he was on the phone.

That was a really big moment for me. Learning to do the mechanical work on a car I cared about, with a friend who knew more than any app ever could.

And every time I hit a decision — every time I had to pick between brand A and brand B, between this tune and that tune, between the downpipe that paired with my setup and the one that didn't — I'd think the same thing:

This is what Cory does. What if Cory was an app?

The two weeks I didn't tell him

I spent two weeks building AutoRev without telling Cory.

This is a pattern for me. I get an idea, I go full head-down, and I'd rather show someone a working prototype than talk about a prototype that might exist. I was in the chair most nights, building a multi-agent system on top of an aftermarket parts database. Agents for parts research. Agents for vehicle fitment. A conversation layer that could actually talk to you like a mechanic would.

At the end of two weeks I texted him: "Hey, let me show you what I built."

He was blown away.

We spent the next couple of weeks refining it together — Cory bringing the domain expertise I couldn't fake, me bringing the product instincts and the AI wiring. That's when AutoRev stopped being my side project and started being ours.

The first version had too many users

Here's what we got wrong in v1.

We were trying to serve three audiences at the same time:

1. People shopping for a sports car. Help me pick what to buy.
2. People maintaining what they already own. Help me figure out what this rattling noise is.
3. People modifying their current car. Help me figure out what to do next with my build.

Every decision we made had to work for all three. Which meant every decision was a compromise. The product kind of worked for everyone and was perfect for no one.

Cory and I spent a while soul-searching. Eventually we made the call: we were going all-in on the modification scene.

Two reasons.

One: the aftermarket scene is massive. If you've never been inside of it, it's easy to underestimate how big this world actually is. Millions of enthusiasts, tens of thousands of shops, entire ecosystems of parts, tunes, and communities organized around making cars better than they left the factory.

Two: it's discretionary money. If someone is about to drop $5,000 on an exhaust and downpipes and a tune, they are already in spend mode. Paying for help making sure they spend that $5,000 right is a no-brainer. It's leverage money — a small amount that protects a bigger amount.

There's a third reason that made the call easy. There are already products that help you shop for a car. There are already products that help you maintain one. There is nothing — and I mean nothing I've found — that gives the guy in his garage at 11pm a real, specific, trustworthy recommendation on what mod to do next, what part to buy, and what it'll actually mean for his build.

That's the gap. And that's where we planted the flag.

What AutoRev actually does

AutoRev is an end-to-end experience for upgrading your vehicle.

It helps you understand what upgrades you can make, in what order, and why. It helps you pick the best parts for your specific car — make, model, year, trim, existing mods, tune level, all of it. And it shows you what each upgrade actually means for performance. Not stacked. Real.

Around that core, we've built community features, events, build logs — the stuff enthusiasts actually want when they're deep in a project.

Running the whole thing is our assistant. We named him Al.

Al is, yes, a pun. Short for AI. Also a great name for the guy in your corner at the auto shop. We have a line we use that's honestly become the soul of the product:

Tony Stark had Jarvis. You've got Al.

Al is embedded throughout the entire app. Quick-action buttons for specific tasks — pull up fitment on this part, compare these two exhausts, tell me what a Stage 2 tune would mean for my car. Or, if you just want to have a conversation, you can talk to Al like you'd text Cory at 11pm.

Under the hood, Al is not one model. He's a coordinated system.

There's a parts research agent. A vehicle fitment agent. Specialized agents for specific decisions that need specific depth. All of them are wired into a purpose-built database of cars, parts, and compatibility data — the stuff that makes the difference between a generic chatbot guessing and a specialist actually knowing.

The reason ChatGPT stacks performance wrong isn't that the model is dumb. It's that the model has no grounded data about how specific parts actually interact on specific engines. Al does. That's the whole game.

The four-week sprint

We went from no product to live in about four weeks.

Start of December: nothing. December 31: AutoRev shipped.

Late nights, whole weekends, Cursor open for most of it, Claude Opus doing a lot of the heavy lifting on the architecture. It was held together with hope in a few places. That's fine. Every v1 is.

Then in late January we did a second launch — rebranded, repositioned, and narrowed to the mod scene. That was the version that started getting real traction. Since then we've been hardening: better data, tighter agent coordination, more vehicles supported, more upgrade paths mapped.

Four weeks of building, one relaunch, and now a product that people actually use. Under the old model, with a developer on the other end of an email thread, this is a 12-month project. Minimum.

What AutoRev taught me about AI products generally

Here's the bigger lesson underneath the RS5 and the exhausts and Al.

Generic AI is great for generic problems. But the real products — the ones people pay for, come back to, and recommend — live in the domains where generic AI falls apart.

Automotive is one. Finance is another. Law. Medicine. Construction. Anything where the details actually matter and a confidently wrong answer isn't just wrong, it's expensive.

The gap between "give me the best cold air intake for an RS5" and "give me the best cold air intake for an RS5 that pairs with my Stage 1 tune, doesn't throw a CEL, and fits the routing I've already got from the downpipes I put in last month" is the entire gap between a chatbot and a product.

Closing that gap takes three things.

Domain data the LLMs don't have. In our case: thousands of parts, thousands of vehicles, compatibility matrices, real performance numbers. You have to build that.

Specialized agents tuned for specific jobs. No one model is good at everything. A parts-research agent, a fitment agent, a performance agent — each doing the thing they're best at, orchestrated together. This is the composability lesson I learned on StoryCraftr and have never stopped using since.

A human who actually lives in the space. This is the one builders most often skip, and it's the most important. You cannot ship AutoRev without Cory. You cannot ship a finance AI product without someone who's sat in a CFO role. You cannot ship a medical AI product without clinicians in the room. The human expert isn't a nice-to-have. They're the thing that keeps the model honest.

That's the real thesis. AI doesn't replace expertise. It scales it.

Still building

AutoRev is live. It's better this month than it was last month. Every week we learn something new about what enthusiasts actually want — and every week Al gets a little smarter about it.

The RS5 is gone. I've moved on to my next build. But the reason AutoRev exists is still the same reason it started: there are a lot of people out there in their garages at 11pm, trying to figure out the right next move, asking the wrong tool the right question.

They shouldn't have to do that.

Tony Stark had Jarvis. You've got Al.