How hardware really gets made — KEEP from one idea to a Kickstarter product in eight months.

Consumer hardware isn’t hard because of the engineering. It’s hard because how it actually gets done is known only to the people who’ve already done it. The datasheets, the dead prototypes, the feature you were proud of and had to cut — none of that shows up in the launch photo.

So I’m making it visible. This is the full backstage of KEEP, my next product: the design, engineering and marketing calls that took it from one idea to a Kickstarter-ready product in eight months. The clean version is a single render. The real version is below, eight steps, the ugly parts included.

01 The unlock: a battery was the dealbreaker, until one email.

For years, Lovebox customers asked for the same thing: let me carry the messages, not just keep them on a shelf at home. The wooden box lives on a desk. People wanted it in a pocket.

The obvious answer was a wearable, and the obvious answer didn’t work. A wearable needs a battery, and nobody recharges a keepsake. A keepsake you have to plug in every night isn’t a keepsake, it’s a chore. Staying connected over Bluetooth or WiFi made it worse — unreliable, power-hungry, exactly the kind of thing that turns a lovely object into a support ticket. So the idea sat there. Wanted for years, blocked on physics.

Then Nicolas — my hardware partner in Hong Kong — sent one email:

The e-ink + NFC tech is interesting. A screen that runs over NFC, with no battery.

No battery. A screen that draws its power from the phone tap that updates it, holds the image with zero energy, and keeps it there until the next message replaces it. The product we’d always dreamed of went from impossible to within reach in the length of one message. That’s how most unlocks actually arrive — not a brainstorm, one person who knows the component you’ve never heard of.

02 Proof of concept: just the electronics, nothing else.

Before designing anything you can hold, you prove the thing that scared you. So the first build had no casing at all — a bare board and an NFC e-ink screen on a desk. One job: confirm a message could land with a simple tap.

The part that would have eaten weeks a few years ago wasn’t on the board — it was the mobile side. To put an image on the screen, the phone has to send the e-ink control board a precise sequence of low-level NFC commands, in exactly the format the controller expects. That format lives in a supplier manual — the NFC FMSC ESL User Development Manual — and it is brutal: APDU commands wrapped in NDEF, a PIN-authentication handshake, load-image and redraw commands addressed by hex codes, a status-word table for every way it can fail. The kind of document written for people who already speak the protocol.

So I handed that PDF to Claude Code and asked it to turn it into a native NFC module for our React Native app — the iOS and Android code that actually talks to the board. It read the manual, wrote the module, and after a few iterations the screen updated on tap.

It worked. The scary part was no longer scary, and only then did we let ourselves design the actual product.

03 The AI-era flip: the render became the brief, not the output.

Here’s the moment that most embodies why I’m writing this blog at all. AI is bridging the gap between creativity and the market — and this step is what that sentence actually looks like on a workbench.

I designed the first real version of KEEP as photoreal renders with Nano Banana 2: a few dollars, a couple of hours, a bamboo keychain reading “I love you” that looked like a finished product shot. Then we handed those renders to our manufacturer of ten years — not as inspiration, as the mechanical brief.

Renders used to be the last step of design — the glossy thing you make after the engineering is done, to sell it. This time the render was the input to the mechanical designer. That’s the AI-era flip.

That inversion is the whole game. The expensive, slow part of hardware design used to be getting from “idea in my head” to “something concrete enough that a manufacturer can quote and tool against.” That used to mean an industrial designer, weeks, and a real budget. Now it’s a prompt and an afternoon — and the gate that used to require capital moved. To be clear, it didn’t disappear: tooling, certification, the first production run still cost real money. But the start no longer does, and starting is where most products die.

04 First prototype: honestly? Not good.

Then the real unit landed in my hands, and the gap between a render and a product showed up immediately. The finish was off. The screen had sync issues. It felt like a prototype, because it was one.

This is the part nobody renders. A render has no constraints. A real product is nothing but constraints — tolerances, materials, assembly, the half-millimetre where two parts meet. Each one forces a compromise, and every compromise shows up in your hand. The render reads “done.” The first unit reads “you have a lot of work left.”

The right response to a rough first prototype isn’t to be discouraged by it. It’s to expect it. One conviction, written on a sticky note: need to iterate.

05 V2: refined in 3D, then finally in hand.

We took the prototype’s problems back into CAD and reworked the whole thing with our mechanical designer — cleaner bamboo, a sharper screen seat, the dimensions locked at 84 × 38 × 6.5 mm. This is the part of hardware that AI doesn’t do for you: someone who has spent years in 3D, thinking about draft angles and wall thickness and how the thing actually comes apart on a line.

When V2 came back, it was a different object. The bamboo read like a product, not a sample. The screen sat flush. For the first time it felt like something I’d want to hang on my own keys.

06 The beta truth: 80% struggled, so we cut the feature we were proud of.

We put V2 in front of real users and watched the metrics instead of our own egos. And the metric was brutal.

We were proud of the 4-color screen. It looked richer, more like the drawings people send. But color e-ink over NFC needed a 30-second sync, held perfectly still — and 80% of beta testers struggled to use it. They moved the phone, the sync failed, the image came through half-rendered. The feature we loved was the feature breaking the product.

So we cut it. Dropping the colors fixed the sync, and — this is the part that’s easy to forget — users didn’t miss them. Black and white, instant. A message that lands the moment you tap, every time, beats a richer image you can’t reliably get. The thing you’re proudest of building is often the thing the user is quietly fighting. Watch them, not your own taste.

07 Validated: “now it just works.”

We shipped the stripped-down, black-and-white, instant version to a second wave of testers. One verdict came back, almost word for word, from enough of them that I trusted it:

Now it just works. Every message lands, instantly, every time.

But “it just works” was only the floor. The feedback that actually told me we had something wasn’t about reliability at all — it was that people found it fun. Testers loved receiving on it: a note or a doodle from someone they love, waiting on the little bamboo screen clipped to their keys. They kept it where they’d catch it during the day. They told us which messages made them smile, and they wanted to know when they could send to more people. That is the part no spec sheet captures — and the part that made me sure KEEP should exist.

08 Now: building the launch — same method as everything above.

We’re preparing the Kickstarter, and I’m building the campaign page the way I build everything now: with Claude Code, writing HTML that generates every visual for the page. No agency, no design studio retainer, marginal spend close to zero.

The pre-Kickstarter landing page, which I wrote up here, came out of the same loop: brief Claude Code, generate, react, ship. The product and the launch page, the same small AI-native cycle.

One honest note to end on: this post was made with AI too. It started as a LinkedIn carousel I put together about KEEP, and I had Claude Code expand that into the long-form story you just read.

What didn’t get automated.

Reading this back, it can sound like AI did it. It didn’t. AI compressed the slow parts; the decisions stayed mine. The ones that actually shaped KEEP:

• The conviction to keep going after a bad first prototype. A render that looks finished and a unit that doesn’t is the exact moment a lot of products quietly stop. Choosing “need to iterate” over “maybe this isn’t it” is judgment, not a prompt.
• Killing the 4-color screen. Cutting the feature we were proudest of, on the evidence of the metrics, against our own taste. No tool tells you to do that. You have to be willing to be wrong in public.
• Knowing which component conversation to chase. Nicolas’s NFC email only became an unlock because we’d spent years stuck on the right problem and recognised the answer when it arrived.

AI gave me the leverage of a team I don’t have. It didn’t give me the calls. Those are still the work — and the part I actually want to keep doing.

What should I write about next?

A few of these steps deserve their own post: turning a brutal NFC ESL manual into a working React Native module with Claude Code, running hardware design through Nano Banana before the mechanical engineer touches CAD, writing a spec for a PCB engineer that doesn’t waste rounds. The full list is on the homepage.

But I’d rather hear from you than guess. If one of these eight steps is where you’re stuck, or there’s a backstage you want me to open up, drop me a line at me@jeangregoire.com. I read every email.

KEEP opens on Kickstarter on June 11, 2026. Follow along at keep.lovebox.love.