An early peek.Solari is a meal-planning app launching later this year.
The build · for the curious

Built from scratch.

I'm building Solari solo. The app, the AI, the brand, even this website. I spent 20 years around some of the best data and AI companies in the world, and I never once wrote the code myself. Then the tools got good enough. This is the story of what happened next.

This page is for the curious. If you just want dinner handled, the rest of the site has you covered.

Three ways in
The tech. How the 4-layer engine actually works. The story. Where it started, all the way to now. The latest The live build roadmap, updated as I go.
The speedrun

The whole thing, fast.

Where it started, and what I'm up to now. Click any moment to jump in.

Where it started

It started with a Tuesday.

5pm. No plan. The fridge half full of stuff that doesn't add up to a meal. And that little hit of dread: what's for dinner tonight?

Every single week, the same exhausting loop. I lived it for years. It's not a huge problem. It's a small one, over and over, that quietly eats your evenings and your energy.

The first version

So I built something. Just for me.

No business plan. No grand vision. I just wanted to fix my own week.

I whipped up a rough app with no-code prompting tools (Replit, Base44, whatever got me there fastest) and started using it. That was the whole thing. An app with an audience of one.

The surprise

Then something weird happened. I kept using it.

I'm the kind of person who falls off after a week. Download something, use it a few days, forget it exists. That's my pattern with basically everything.

Not this one. Week one. Week two. Week three. Week four. I kept showing up. I kept adding to it. I could not put it down.

That had never happened to me before. That got my attention.

The proof

What actually changed in my week.

Not in theory. In my real life, that first month.

Turns out

It wasn't just me.

I started asking friends. Do you have this problem too? Would you actually want something like this?

The answer came back fast and loud. A resounding hell yeah. Okay, whoa. This wasn't just my little hack anymore. It was a product.

All in

So I quit my job.

Twenty years in tech, and I walked away to build this full time. Not because it was safe. Because I couldn't not.

What I brought with me

I was never the engineer. But I was in the room.

I'm not a developer. I never wrote production code. But I spent two decades right next to some of the most interesting data and AI work anywhere, and I understood it cold. I just couldn't build it myself.

Eight years at Force Management gave me a front-row seat to how 40-plus tech companies built their go-to-market, their product, and their exit. So I got to see it done 40 times over. Then I went and helped build three companies from the ground up myself.

Early at MongoDB

Where I learned data at a deep level. Structured versus unstructured. Data in motion versus data at rest. What real infrastructure actually takes. The plumbing under everything.

Gretel, acquired by NVIDIA

Early to agentic AI and synthetic data. One of the reasons NVIDIA acquired us was our work on AI as a judge. Hold that thought, it comes back in a second.

Roboflow, the last two years

Deep in computer vision AI models. Watching, up close, what becomes possible when machines actually understand the world.

The full resume, if you're into that: Jeremy Powers on LinkedIn.

How I build

From vibe coding to a team of agents.

I'm still learning daily, but here's how things evolved over the build so far.

First, I vibe-coded the thing. Pure prompting, off-the-shelf tools like Replit and Base44, just describing what I wanted until it worked. Got me a real web app and proof the idea had legs.

Then the brand and this site were a totally different kit. Claude Cowork and Claude Design, plus a pile of image models like Gemini image, to get the logo, the look, and every screen right. I was a psycho Claude Design power-user, spending hundreds. Then I ran the whole thing through Gary Tan's G-stack: a CEO review, an engineering deep-dive, a design review. Startup rigor, one-person shop.

And now I'm getting into agentic engineering. Software 3.0 if you want a label. I honestly don't care what it's called. In practice it means I work with agents that each own a specific job. A team of engineers and marketers, except they're agents and I'm directing them. That's how I'm building the next big thing: a genuinely strong native iOS app.

Have I nailed agentic engineering? Hell no. I'm directionally there, not done. But it's a real step past vibe coding, and it's the most fun I've ever had building anything.

A verification result reading PASS 22 / 22 / 34: 22 tables created, 22 with row-level security on with not a single table left exposed, 34 access policies in place. A note reads: for a first time doing database work, you just stood up the entire backend foundation correctly.
Week 1, locked. My first time ever doing database work. 22 tables, every one secured.
Under the hood

Most AI apps are a thin wrapper. This one isn't.

This is the part I'm proudest of. Most "AI apps" are a thin wrapper around a single model call. Solari is the opposite. The AI does the smallest, safest part. Plain, deterministic code does the heavy lifting.

Drafting your week runs four layers. The AI only touches the last two.

The AI never decides what's safe or what's feasible. Code does. The AI only adds taste on top of an answer that's already correct.

A lot of builders reach for an LLM because it's the easy button. I used plain code wherever it's actually the better tool, and saved the AI for the two jobs it's genuinely best at. On top of that I'm wiring in real APIs, like Instacart, so the plan flows straight into a cart.

A deeper teardown is coming as I open more of this up.

An honest moment

Some of this is way harder than it looks.

Here's the one that humbled me. Getting a recipe into Solari from wherever you found it. You paste a link or share a reel, and it just works. That's the promise. The reality behind that one sentence has eaten more time than almost anything else I've built.

I thought it was a checkbox. It turned out to be an iceberg. Every place people save recipes is messy in its own special way.

So instead of one tidy importer, there's a ladder. Try the clean, free path first. Then escalate: read the page's structured data, follow the bio link and hunt down the real recipe, pull the full untruncated caption, transcribe the spoken audio when there's no text at all. And if every rung fails, hand you a half-filled card to finish, never a dead end. Plus residential proxies for the sites that wall us out, and guardrails so a bad link can't do something nasty.

Is it done? Not even close. It's live for a lot of cases, and when it lands it genuinely feels like magic. But there are still creators and formats I can't read yet. "Never make you type it in" is the bar, and I haven't fully cleared it. This one is going to keep getting work for a while, and I'd rather tell you that than pretend it's solved.

The receipts

By the numbers.

20
years in tech. Never wrote the code.
40+
tech companies I watched build, sell, and exit. Eight years at Force Management.
3
I helped build in the trenches. MongoDB, Gretel, Roboflow.
4
layer plan engine. Only 2 layers touch AI.
2
model calls to plan your whole week. The rest is plain code.
0
critical gaps before a line of app code. 25 decisions logged first.
Why I'm sharing this

Why I'm telling you all this.

Because it's a blast. Genuinely. I own this problem, I understand this space, and I get to build at the edge of what's possible, solo, every single day. That's rare, and I don't take it for granted.

I'm not running a daily blog. But I'll keep notes here as I go, for the people who find this kind of thing as fun as I do.

See the live build roadmap → Follow the build on LinkedIn → Know a family who'd love it? Send them here →
Jeremy · building Solari, from scratch.