https://www.simonberens.com/p/lessons-learned-shipping-500-u...

HN thread: https://news.ycombinator.com/item?id=46848876

... But also that needs to be an official measurement, and if you're sloppy you can measure it wrong during your own testing and then be surprised.

Not only that but every single update you push requires paying $$$$$ to get certified. It is freeing that the cost of releasing a fix or improvement to a product is $0.

In the past I considered selling a custom adapter to hobbyists, but all the certification costs just did not make it a viable option to pursue. I didn't want to up front thousands of dollars to make like $100 if I was lucky.

Back in the glorious past, the iOS App Store either didn’t allow apps larger than 100 MiB, or at least forbade them from being downloaded over mobile connections. There’s an old blog post/Twitter thread you can find on Hacker News where an Uber engineer describes the challenges they faced in trying to keep their app under that limit. They were also rewriting the app in Swift, so some compiler patches were necessary.

yeah, and looking at their board I see they use one of "ESP32-WROOM" modules, which has the wifi & blutooth & antennae (pre-certified by Espressif for FCC, CE so itself doesn't need testing) & flash & processor (which even has its own RC oscillator), which takes care of a whole ton of difficult stuff so all the designer needs to do is plop that module down and power it up and wire some SPI/I2C peripherals which I'm guessing don't need super fast clocks or electrical constraints. Without something like those ESP32 modules, this would be much more difficult.

Has that happened to you?

How did you find customers?

How did you get your hardware design verified?

1) Does device record ALL MIDI sent to it? (Including e.g. SYSEX and non-registered parameters?

2) I edit MIDI on a Desktop (NOT a phone). a) Are sessions recorded into separate files? Named and/or dated? b) Can I inspect & load sessions from the card?

I was originally planning to do 4000 units as my first run. Boy I'm glad I started to run out of money and cut it way back. That would have been way too much as a first run.

Can you expand? What counterfeits do you see (literally someone selling boxes labelled jamcorder? How do you handle that? Lots of scanning amazon and sending cease desist letter?

That doesn't mean it is completely impossible, though. Fab2 (formerly Atomic Semi, of Jim Keller and Sam Zeloof fame) is trying to do exactly this by aiming at making small-scale production affordable. The trick is to replace the whole mask-based lithography with electron beam scanning: way slower, but no mask costs.

But they are now targeting selling mini-fabs to R&D labs rather than offering an affordable ASIC-as-a-Service. Let's hope one of their customers makes it available to the public!

These niche hobby products are usually reasonably safe because they rely on community network effects, becoming the darling of a random subreddit or discord group and people aren't going to save $20 to get some device of dubious origin.

I wrote the software, and my friend designed the PCB for the digital side. Basically I had to be a decent C programmer, and read the datasheet for the MCU we were using (which was way more work than using something like an ESP, but still). My friend had to lay stuff out on the PCB and place secondary components according to various application notes.

Overall not simple, but not rocket science.

On the other hand, the guys doing the analog daughterboard were industry veterans with decades of experience. They had a million concerns, from signal integrity, to voltage protection, back-EMF, etc. I was not even near the ballpark of that skill level.

Later on, I worked on more complex HW projects, stuff with FPGAs, DRAM integration and 8 layer PCBs, where you had to run full electromagnetic simulations because when you run at 1+ GHz, suddenly signal propagation is not instant, and traces behave like antennas.

I also did software there (a smaller slice of it this time), but there was an army of engineers working on the whole project which took a year, and not a second of it was spent in vain.

If I ever get good at piano it'll be handy

It is mentioned elsewhere here that a lot of time was burned on unused firmware features. It takes surprising discipline to cut stuff and still deliver.

Congrats though, making a half-million a year business is a massive accomplishment! The headline just irked me a bit, haha (but, it got my attention, so I guess you win there)

Low product quantity and low quality/fake products resold on amazon from AliExpress are the biggest roadblocks at this point. The juice isn't worth the squeeze.

I suspect we will see a reversal on this. LLMs are flat out worse than humans at programming (yes, even your latest favorite frontier model). I am faster (once you factor in the time fixing its inevitable mistakes) and better at programming than Claude is, and I'm not a particularly exceptional programmer. People say "they aren't going away", but I disagree with that - already we've seen people disenchanted with the lack of results compared to what they pay, and that is only going to grow once the AI companies raise prices in an attempt to turn a profit rather than lock people in. Eventually the industry will wake up and stop drinking the kool-aid, and what matters will once again be the results you deliver, not the tools by which you deliver them. But it's hard to say when that will happen, and it's going to be rough in the meantime.

First, so far, there have been 0 failures in the field. My email address is on the website, the app, and the first page of the instruction sheet, so I'd probably know if there were.

But since it's fun, I'll address some of these :) I'm not flying by the seat of my pants here as much as you expect.

- The ESP32-S3 includes ±2 kV ESD, which is respectable. Adding more is a bit precautionary IMO, but I did add it in some places where the user is more likely to touch metal (like the USB ports). The sdcard's metal housing is not exposed.

- The AMS1117 is just more available than the alternatives. Again, no failures yet so seems like the right tradeoff was made here.

- Supercaps don't last long enough for my liking. A CR2032 can go for 10+ years. A supercap, only weeks. I want correct dates even if the device is left unpowered for months.

- The 0603 47ohm resistor is a good callout. Worth beefing that up. But at 3.3v and 232 mW it's not a fire hazard. It will fail open.

- USB-B for power was very intentional. It's a great firm connector and more importantly prevents users from plugging it into their piano wrong. Many of my customers are in their 70s or older.

- USB-A for data was made for manufacturing simplicity, but again, USB-C is too complicated for many of my users. I'd only switch to USB-C if both ports could be used interchangeably, but they can't be / it's not worth the complexity.

- As for the software moat, time will tell. You can't stave off competition forever!

- The device is FCC and CE compliant.

Again, I actually do appreciate your critiques!

(Although I spy an SD card slot, a favorite of the "hardware isn't hard" faction that will explode in their face)

Any tips on how or where I could get my design reviewed?

99.99% of cheap Chinese products you buy on Amazon don't actually have this testing done at all. So by doing any testing you're already way above them.

But why do you spell MIDI in lowercase there?

> plugs into your piano via midi

Maybe I'll write an article about that.

Electronics, PCBs, CAD, injection molding was all new to me.

Here's a good example: https://duckduckgo.com/?q=brick+smartphone+blocker&ia=web

Never understood this, was the "most of the fun of programming" the literal typing for you? For me it was about solving problems and building things, and I get to do that more now than before, and without all the literal typing.

At this point I have the only stock of this part in the world, and it's more than a lifetime supply, but otherwise I'd have been faced with a re-design and limited options.

For another of my products, during the pandemic, the part I was using became unavailable, and I had to re-design my board to accommodate a part of the same type in a different package. Then I couldn't buy either part, but found an odd lot at a Texas Instruments factory in China, which I bought.

Dealing with shortage and obsolescence issues is a constant battle. There have been crises in the industry caused by plants catching on fire.

Apple UX used to be good, not anymore sadly

https://www.uber.com/us/en/blog/how-uber-deals-with-large-io...

Comments about a Twitter thread talking about the same thing: https://news.ycombinator.com/item?id=25373462

I did this as an experiment on my current project and it found a bunch of minor, non-critical, but completely valid problems with my deisgn.

The real answer: I got tired of deciding what should and shouldn't be capitalized. Sounds easy but it gets ambiguous fast.

I'll revisit this decision at some point. Capital letters do make reading easier.

CE is all about self certification. The lab testing gives you the confidence to do this.

It can be expensive - but Chinese test labs are very affordable and do a really professional job.

There is now quite a good EU based service - https://early.smander.com/ I used them to get some advice early on (disclaimer, I know the founder).

[1] https://www.kicad.org/community/chat/ [2] https://www.reddit.com/r/PrintedCircuitBoard/

Yeah, but those Chinese dropshippers don't have significant local presence. If they are caught doing shady stuff they'll just let their $100 letterbox company go bankrupt and start over.

You, on the other hand, would actually have to pay whatever fines some regulator charges you with, because you are physically present in a jurisdiction they can apply power to. This makes self-certification a bit of a footgun: you essentially pinky promise that you fully understood all the relevant rules and regulations, and otherwise pay what in the worst case might be a hundred-thousand-euro fine. Quite a bit of a gamble!

The main saving grace is of course that the authorities are way too busy to investigate some random guy selling a handful of devices from his kitchen. But get too successful, and some competitor might just decide to give them a little hint on exactly what part you missed because you didn't even know you had to comply with it.

And, well -- that's certainly good.

But the FCC rules cover more than just Wifi and BT, and the whole rest of the widget also needs to comply with them. Of particular interest is FCC Part 15 Subpart B, which covers unintentional radiators. The device would need to comply with Part 15 Subpart B even if it had nothing to do with Wifi or BT.

Putting one part that is certified for one issue into a device doesn't extend a magical blanket of compliance to everything else that the whole of the widget contains.

Does the SPI or I2C bus (or anything else) leak RF beyond allowable limits? Eh, who knows. I certainly don't know. It's easy to say words like "it should be fine," but without testing then nobody actually knows.

The CE mark, meanwhile: That mark means exactly nothing in the US, where the author is located and the devices are sold from. The United States has zero regulatory recognition of any of the processes and/or tests that CE may cover. People often confuse this issue, wherein: They don't think it be like it is, but it do.

Just because you use a certified module doesn't give you a free pass.

I don't disagree. It's important for us to recognize that we stand on the shoulders of giants, but I also think it's a little silly of a critique of someone happy about it being easier than he expected. It's true that hardware would be much harder if you had to do it all yourself, but is that meaningful? The wonders if the modern supply chain is that you don't need to do it all yourself.

But for providing power over USB-C, you generally need a dedicated IC to handle VBUS switching. It wasn't worth it.

Also, once you ship a consumer product with a USB port on it, you'll realize a lot of people really don't understand how USB works, and USB-C doubly so. Square holes for square pegs goes a long way.

The firmware is about 110K lines: MIDI recording + streaming, USB, Bluetooth, Wi-Fi, web APIs, storage, OTA updates, diagnostics, security, peripherals, provisioning, and an embedded web interface.

There’s a lot required to make a full-featured product.

This is (literally) none of my business, but would you consider a cut down or basic open source version (hardware and software, with open software running in the proper hardware). Yes it'll be cloned by factories, and yes it will cut into sales no matter what proponents claim. (And yes it's probably more work than I imagine it is).

But also yes, it'll be genuine open hardware and might out live your interest in the project.

Beyond the typing though, I find enjoyment in perfecting the structure of the code, thinking deeply about how the pieces fit together. I enjoy the process of iteratively refining the code so that it clearly represents the program using the optimal level of abstraction and minimizes superfluous code/syntax. I find that using an LLM eliminates all of this and produces code that lacks all of the elegance and readability that I aspire to in my own code.

The problem at least I have is... when something goes wrong, often enough even you as the original author / LLM slave-driver have no idea what is going on where, and completely forget trying to dig into someone else's pile of vibe-coded garbage. For every single problem you now are completely beholden to the AI agent.

Your take is so reductionist it's hard to believe you actually like to solve problems.

Some people fall to that temptation instead of using LLM to just do the menial stuff and do the fun stuff themselves.

You hear about firms that go over certain thresholds, get noticed and then find out that their production facility is no longer "small" or exempt for example.

Some amount of defensiveness in planning and design can mitigate but I'm interested to know if there are thresholds I didn't know about.

To be fair, it's unintentional radiator testing that's required, rather than everything the module had to go through as an intentional radiator. So, it is a "free pass" for nearly all of it, and a much cheaper one, since you don't have to use a costly FCC accredited lab.

LLMs move me away from the implementation, which is more a managerial role, something I personally have zero interest in.

I think there's a lot of people talking past each other here. Sourcing parts internationally, preventing counterfeiting, slow iteration time, inventory management...all this logistical stuff is hard enough that the conventional wisdom in the software business world is to never attempt it. You're here presenting a counterexample and that's wonderful. The people criticizing aren't talking about this though, they are talking about whether logging digital inputs on an esp32 counts as a real hardware project and they act as if you've insulted their expertise in [signal processing/antenna design/insert specialty here]. You didn't do that, most of us know you didn't do that, but the word "easy" sets a lot of engineers off.

Just ignore the noise and keep writing great articles. My personal request, if I may - I'd like to hear how you did injection molding since I always assumed that was out of reach for small business.

If you want a Free work-alike, why don't /you/ put in the work and make it happen?

In the past when my junior engineers were fighting with a bug all week, I always told them the dopamine hit of solving the bug is what keeps you coming back to work for 20 years. Half joking, but the dopamine hit of from the satisfaction getting complex working software in a weekend is like crack to me, we'll see how that plays out long term :)

I'm sorry if I gave the impression I think programming is just typing. For me "programming" is the whole activity of creating software with code, you can do programming without the literal action of typing.

> Writing the code and thinking about the exact logic you want to apply and all its implications are part of the same problem solving process in my mind.

That's really interesting perspective to me, as these two activities for me are two completely things, and "thinking about the exact logic" happens way before I actually write any code at all, sometimes weeks/months even.

> putting things together in code is a very satisfying and IMHO is the actual building part of the job.

I agree that it's satisfying, but I guess "putting things together in code" still for me is more about being able to mash two concepts together in an elegant way, and none of that actually requires actual typing still.

> LLMs move me away from the implementation, which is more a managerial role, something I personally have zero interest in.

This is also interesting, because it feels like LLMs give me the opportunity to do more implementation, and not get stuck in the minutia so much, but rather focus on how the design and architecture actually fits together. Refactoring no longer is "I want to but also I don't want to spend the time nor get stuck in more rabbit-holes", so being able to focus on architecture/design more feels like it lets me care more about the implementation than before.

1. https://www.europarl.europa.eu/doceo/document/P-6-2007-5938-...

Nearly every component gets tested: sdcard, clock, battery, anti-counterfeit IC, led, buzzer, the tactile-switch!

And they’ll make the molds in China anyway. Chinese molders are hungry — they will work with low volumes and offer a bunch of integrated services. But if you don’t know what you’re doing you will end up with bad parts over schedule.

Easiest way to get started is to use ProtoLabs. They charge lower prices on molds in exchange for higher part cost, are way faster than China, and have way better customer service.

If there are no more app updates, presumably there are no more firmware updates either, so the flashing part probably doesn’t matter (and you can probably flash the last published image to an authentic device by other ESP32-y means if you really need to).

As to the day-to-day: Pretty sure the hardware still records regular ol’ MIDI files to the SD card. The app seems incredibly well-thought-out and ergonomic but, at the end of the day, nonessential to the device’s core function.

Somebody could presumably build and sell their own “MIDI Recorder Loop Nonstop Device Digital Piano Upgraded New 100%,” but it wouldn't be a Jamcorder, you couldn’t flash Jamcorder firmware updates to it just by slapping an ESP32 in a case, and the official Jamcorder companion app would not talk to it. Which… fair enough, right?

For that matter it wouldn’t surprise me if somebody could write their own client for the real Jamcorder, if they felt so inclined… but with the official stack so absolutely perfect from a design perspective, why would they?

Serious question: why?

I figure that the factory is exactly the adversary in the whole threat model? So why give them the keys to the castle so they can moonlight their own genuine batch?

I suspect that 110k number might include the ESP-IDF libraries as well as his own code.

I've written a fairly feature-complete handheld device based on the esp32-s2 and the codebase is about 9k sloc of the C I've written, but that number explodes to almost 40k sloc if I include the esp-idf's imported libraries.

With code, I'm not planning everything down to the finest detail before I start writing. I know what I ultimately want my code to do but there are many unknowns I hadn't considered. Coding with an LLM is different depending on how you use it. When I'm using code completion, that's usually just helping me prevent syntax errors or to write an efficient snippet of code without having to check docs. But when I tell Claude "I want a web app that does xyz" and it creates it whole-cloth, that takes away the challenge of learning something new and building. I'm not a web developer and anything I could create is going to take me a long time to make (and it probably won't work well or look good) but Claude cranks it out with ease. I get to play software project manager and let the expert do the technical work. It's kinda fun being able to quickly get resulrs. Rather than debugging individual lines of code and thinking about the best way to implement code, I get to test features and think about better ways for the overall system to operate.

It's more the opposite. The app requires a genuine device.

Maybe they just like doing it the old fashioned way. You can both exist.

Unless you’re getting a commission to sell LLMs to people I struggle to understand what you’re trying to achieve.

A local-only device with no telemetry can probably skip some of those

The value you have created is actually more in the App than in the hardware.

As a maker of small devices (mostly open source) including MIDI - and putting aside commercial aspects like injection molding - I could recreate the hardware side of this project without much issue. The necessary components are likely already in this room with me. (I won't - I'm not in the business of copying things that people don't want copied)

But I'm not a software engineer, and therefore could not produce software of the scope and quality that you have, which is where the unique functionality of your device sits high above the relatively simple "receive MIDI data from MIDI devices" principle of the hardware by itself.

It's not bothersome, I'm not bothered by it at all, everyone gets (in my mind) to do things that makes sense to them. When I write comments here disagreeing with others, it's more from a "Please explain further your perspective so I can understand it, which hopefully I trigger by challenging it" view than anything else, not "You should do as I am doing things".

> Maybe they just like doing it the old fashioned way. You can both exist.

Completely fair if that is the perspective, just to re-iterate, I'm not trying to convince anyone of anything, only to understand other perspectives better.

You probably aren't going to do this whole rigmarole for small runs of devices like in the OP.