Hacker News Reader

Comment by gregsadetsky | Hacker News Reader

Was it working / were you making money?

Automated handwritten notes companies seem like they render text using a handwritten font and then potentially pass it off (via API) to a “print anything and ship it” company — is that an accurate statement?

Is there a clear advantage to building and maintaining the hardware / plotting pen portion? i.e. isn’t that something you could do by randomizing font parameters?

Also, it does seem like a pretty crowded space - was that ever a problem? Googling “handwritten notes automation” brings up a lot of companies/ads.

Thanks. Really curious about this!

View on HN

I feel like I've been able to detect these, e.g. examine the writing and note that each letter is exactly the same every time it appears -- real people don't write that precisely. Though I'm sure that some random variation would not be difficult to introduce.

Bottom line it's like anything else -- if I don't know the person who sent it to me, I don't trust it. Automated handwriting is just more big-tech destruction of yet another facet of human to human communications.

I was about three years behind wallawe’s reply.

The technical parts are fun…the core thing for us was that it was actually written by a real pen. The difference between a print and real writing is obvious to most people, so we wanted it to be believable.

Our system used a set of handwritten fonts, along with a bunch of “fuzzing” of line height, line direction, and character swap outs. This created an SVG path that a robot used to trace with a pen.

The hardware and margins were what killed us. As wallawe noted, there’s still a lot of manual labor that goes into a system like this, even if the writing is automated. Envelope stuffing and card-envelope matching are big problems.

Its not a super competitive market yet, but the margins on what could be charged per card were pretty low, especially factoring in paper costs and labor costs.

> Our system used a set of handwritten fonts

What betrays handwritten fonts is nobody writes "a" the same way every time. I get that you're addressing this with fuzzing, but that's not enough. The pressure put on a pen also varies, people make mistakes and cross them out, there are ink blotches, etc.

There were enough character alternates and special compound ligatures that it’s really hard to tell unless you know it’s machine-produced.

I think you would be surprised at how realistic it can be.

Perhaps. Just yesterday I received a handwritten note from a business. The varying pressure on the pen was clear if you ran a finger over it, as some areas were more deeply indented than others. The line thickness varied, sometimes trailing off into nothing. No two letters were the same, some were illegible in isolation. The lines would even overlap. Every ligature was different.

I've never seen a machine handwrite that way.

I'm sure they'll get better, but there's a long way to go.

I replied to parent, but we built the company that shows up #3 for that query (at least on my browser) - handwrite.io. I felt that there would be a massive audience for the API so worked on building that out in the early stages while my cofounder built the hardware. It's a super costly process, requires a lot of human touchpoints even when you've automated a lot of the machinery (e.g. sealing and stamping cards). There was a decent amount of competition but even more demand. We had a couple of good sales guys working strictly on commission that got us to 40k MRR ish in a relatively short timespan but the margins were horrible. In the end we burnt out and sold to a client.

Thanks a lot for sharing, really interesting.

I edited my comment so you may not have seen my additional question: why build out hardware vs. outsource the printing / mailing?

i.e. with enough randomness in the font/text/letter generation, was it still obvious to tell what was “handwritten” by a plotter vs printed?

And re: margins, you could have raised prices but that would have made you not competitive?

It is possible for a plotter using a real pen to make a card which is indistinguishable from a truly handwritten card.

People are rightfully squeamish about sending out cards which “look” handwritten but aren’t. Imagine a customer realizes that you are faking it…the consequences to reputation could be substantial.

That said, I personally find printed cards using handwritten-style font to be insulting, so that was a no-go for us.

We considered using a third party fulfillment service and essentially charging a fee on top of that, but the returns would be microscopic for the effort.

it depends on who you ask and how hard they look, but in my experience yes. i did numerous non-scientific experiments of just showing people a card and asking them if they thought it was handwritten or done by a robot. sometimes i would show them an actual handwritten card, sometimes the robotic one and for ours at least, they had a tough time telling them apart.

edit: i can't speak to the technology the company uses now, i believe they switched away from our tech to something more scalable so it's probably not as good now.

> why build out hardware vs. outsource the printing / mailing?

we wanted the handwriting to feel as real as possible. to that end, my cofounder wrote an algo that varied each character slightly, the robots held real pens, etc. We couldn't find anyone to outsource the fulfillment to that matched this level of authenticity. Some competitors used lob.com and simply print an image of handwriting which actually can work pretty well, but is definitely recognizable if you look closely.

> re: margins, you could have raised prices but that would have made you not competitive?

This was a big point of contention between me and my cofounder. We raise them towards the end some but not enough. After selling, the company that bought us raised them substantially and are doing very well from what I understand. We should have done this early on, but focused too much on winning the deal which is what burnt us out in the end (too much business too quick, lots of sleepless nights).

How can you spot the difference between a robot written letter and a printed one?

And what about volumes, is each letter has different text? Should have slightly different writing?

Or are there some volumes in the system?

There are variations in pressure that are really easy to recognize.

Not sure I understand the second question.

Edit: Do you sometimes print the same letter(or a templatized letter) 100/1000 times ?

This would have been a good optimization but we did not. Every letter generated was unique even if it was a templatized piece.

Do you sometimes print the same letter(or a templatized letter) 100/1000 times ?

Reply on news.ycombinator.com