Mechanical Autopen Simulation

2026-02-16T00:00:00+00:00

Or how I nerd sniped myself and ended up reverse engineering Lyndon B. Johnson’s signature</p>

The Inspiration</h2>
The other evening I was killing time browsing a government surplus auction site. I had filtered it to relatively nearby auctions that were ending in the next 24 hours. Among the usual broken-down school buses and old computers, something caught my eye. A listing titled “One (1) Sigtech800 Automated Signature Machine Table”</a>, even from the small picture on the search page it immediately drew my attention. The vintage metal mechanical mechanism taking up the left third of a fairly sizable desk was more than compelling, I am a sucker for vintage mechanical devices.</p>
</p>
When I first found the listing, there was only half an hour left in the auction and only a single bid of $5. It took a lot of willpower to not place a bid myself, but in the end my better judgment won out and I resisted buying this hulking machine I certainly don’t have the room for. My curiosity was, however, piqued, and I descended into what ended up being hours of researching, then simulating these old machines.</p>

Background & Research</h2>
I’ll start with a quick primer for anyone who hasn’t had the misfortune of becoming aware of what an Autopen is because of American politics. An Autopen is a device that can write out an individual’s signature. The general class of devices is referred to as “signing machines”, but much like Kleenex or Band-Aid, the Autopen brand name has become synonymous with the whole class. Signing machines do exactly what their name implies, they allow an individual to create a reusable representation of their signature, which can in the future be used in the signing machine to create a limitless amount of near-perfect examples of their signature on command. They have been used widely by politicians, famous individuals and others who need to sign a lot of paperwork since the 1950s.</p>

Hopefully you never had to see this before. But you can clearly make out the Autopen specific 4-bar linkage</figcaption> </figure>
Modern Autopens are entirely computer controlled. They are essentially robotic devices that can take a digital representation of a signature and using electronic motors move a pen to render that signature onto a piece of paper. Their mechanical design does share some fairly obvious lineage with their pre-computer ancestors, however. These computer-controlled Autopens were first released into the market in the 1980s.</p>
The Autopen I came across, the Model 80, predates the computerized designs. These mechanical Autopens instead use a physical representation of the user’s signature. The Autopen company refers to them as a matrix</strong> or template</strong>. Each matrix stores a single signature and can be loaded in or out of the machine to allow multiple users for a machine and to allow a signature to be stored safely out of the way when it is not in authorized use.</p>
</p>
There seems to be very little written or recorded about these devices and how they work. At least that I could find searching the internet. The best resource I was able to find was the official website of Damilic</a>, the company that owns the Autopen brand. Their site has a few short “Info” and “FAQ” pages that give a couple of paragraphs of explanation for some of the company’s history and legacy machines. Somewhere, in one of those pages was a link to the absolute best resource I was able to find through this whole project: a scanned PDF that Damilic had posted of a 1970s article from the Journal of Forensic Sciences</a>. The article spans 7 pages and goes into a fair bit of detail on how the mechanical Autopens work, as well as having a labelled close-up shot of the primary mechanism of the machine.</p>
Just by staring at the mechanism, I could see how the two pins that ride along the matrix could provide the full two degrees of freedom needed for a signature. If both of the pin arms move in towards each other, the arms would swing outward and the pen would move towards the rest of the machine. Likewise, if the pins move outwards, the pen would be pushed away from the machine. Finally, if both pins move in the same direction, the pen would move in the opposite direction. None of these kinematics are simple though. Being built around the intersection of circles, there is a limit to how far my intuition could take me. So, while the article went a long way towards helping me understand the device, I wanted to find a way to build a deeper intuition for it.</p>
</p>
One other amazing resource I came across was a high-resolution picture of President Lyndon B. Johnson’s signature matrix from his time in office made available by the National Archives.</p>
</p>
The Simulation</h2>
Having gathered enough information to get a rough understanding of the device and then seeing the President Johnson matrix, I began to form an idea. I decided I’d like to try and digitally simulate one of these mechanical Autopens, so that I can play with it and watch it and try to build an intuition for how they work. And finally I wanted to see if I could reproduce President Johnson’s signature from the picture of his matrix. I did not have very high hopes I’d actually pull that second part off.</p>
I started by using the picture in the journal article to draw a sketch of the 4-bar linkage that makes up the heart of the machine and assign rough dimensions to all the pieces.</p>
</p>
LLM Coding Adventures</h3>
I won’t be going into too much detail on the code for my simulations. I used Anthropic’s Claude Code to help with much of the development and in many ways don’t consider the code fully mine. It is not a codebase I would feel comfortable releasing under my name without a good amount of refactoring. With all that said, I would not have been able to get nearly as far in this work as quickly as I did without the use of Claude or a similar LLM coding tool. I will generally say that Claude was not very useful for the physical kinematics, but did much of the heavy lifting for the computer vision, UI creation, and general script boilerplate.</p>
I started by having Claude make a simple visualization of the 4-bar linkage interacting with a matrix. At first this was a random matrix that Claude had created that did not map to any actual signature. Having the visualization was critical for me, it was a very quick sanity check that the physics were being modeled correctly. Once the model seemed to be working for this random test template, I decided to move on to the President Johnson matrix.</p>
Computer Vision Detour</h3>
To simulate the matrix, I needed a tool that could be run on the source image and create an output file with the curve information required to run the simulation and visualization. There ended up being 3 main concerns to deal with:</p>

Extracting the curves of the matrix edges</li>
Extracting the pen pressure/lift information from the top ridge of the matrix</li>
Determining the center of rotation of the matrix</li> </ul>
Matrix Edges</h4>
The image I was using is of quite high quality and the matrix has a high contrast against its background. This all makes the edge extraction a fairly simple computer vision task. Claude managed to get it mostly correct on its first try. If I am honest, once I confirmed that the edge extraction looked visually correct, I never dug too deep into what techniques it used to accomplish it.</p>
</p>
Pen Lift</h4>
The shape of the top ridge of the matrix controls the downward pressure of the pen. As the ridge rises, it presses down the pen like a see-saw. Just looking at the source image, it was clear to me that it would be much harder to extract programmatically. For that reason, I didn’t bother to ask Claude to do it. Instead, I had it create a UI where I could click at each point the ridge starts or ends. The tool then extracts that alongside the curves of the matrix edges and saves it to a file for use in the simulation.</p>
</p>
Center of Rotation</h4>
As I iterated on the extraction and resulting simulation, it quickly became clear that having a precise center of rotation is critical for getting a good output. Originally I had Claude determine the center of the curves it extracted, and then I manually provided an offset from that in the form of an angle and magnitude. I even had Claude create a script to sweep a large set of angles and offset magnitudes, producing ~100 images. This managed to get a very reasonable output but seemed unacceptably brittle.</p>
Finally, while looking back at the images of the matrix and the Model 80, I noticed that there were 5 holes in the matrix that are seated into pegs on the turntable of the Autopen. From images of the Autopen device, it was clear that the turntable center was in the exact center of these pegs. With this insight I was able to have Claude add a step to the extraction script where I could click on 3+ peg holes on the matrix and it would use that to describe a circle and from that circle determine the center of rotation.</p>
</p>
Extraction Results</h4>
With all three of those handled, this is what the output looks like visualized. In actuality it is all stored in a binary file that can be passed to the simulator or visualizer. You can see in green the pen up and down movement. It is clear to see that the red and blue traces of the matrix have a global shift over the course of the matrix relative to its center. This is what encodes the left-to-right movement of the signature being written out. The global shift moves the pen left to right while the local variations are encoding the movement for each character of the signature. You can also make out at the very end where the curves move sharply while the pen is lifted only to drop back down momentarily for the “.” of the “B”.</p>
</p>
Results</h2>
</p>
In the end, I was very pleased with the results I got from the President Johnson matrix image. Below you can see a side-by-side of my results and listing image for an eBay listing for an Autopen LBJ signature</a>. The issue that stands out the most to me is the misplacement of the “.” after the “B”. From my prior experimentation, I know this can be resolved by manually shifting the center of rotation of the matrix just slightly. If I had to guess, it is a consequence of the source being a real picture from a camera and therefore not being a perfectly orthographic projection, which is an implicit assumption of the extraction tool. The image could also be slightly off of perfectly flat relative to the matrix.</p>
If I find some more motivation or this generates significant interest, there are some next steps I’d like to take in this vein of work. I’d love to recreate the simulator in TypeScript and create a website that allows a visitor to use a touch input or mouse to create a signature and then have them be created as a virtual matrix. That virtual matrix along with the LBJ matrix and maybe any other famous matrices I manage to find could then be simulated and visualized. Finally, I’d really love to be able to have the site export those templates as STL or DXF files to be 3D printed or laser cut (presumably at a smaller scale than the 2 ft across of the originals). I could then design a 3D printable Autopen type device. All of these feel very doable but would also require quite a bit of time investment. It would make a very interesting conversation piece and engineering demonstration though. It would kill it at a science or engineering museum.</p>

Simulation Results (left) vs Real Autopen Signature (right)</figcaption> </figure>

Internal Mic Fix for Lenovo T14s Gen 4 w/ AMD on Ubuntu 24.04

2024-04-26T00:00:00+00:00

I recently purchased a Lenovo T14s Gen 4 with the Amd Ryzen 7840u and loaded the brand new LTS version of Ubuntu on it, 24.04. To my dismay the internal microphone was not recognized at all in any of the sound settings. Finding a solution took a couple hours of reading logs and Googling so I’m putting that knowledge here to helpfully help anyone else in the same situation.</p>

Molded Silicone Parts for a Vintage Honda CJ360T

2024-02-25T00:00:00+00:00

Background</h2>
I own several motorcycles, and in my pursuit of novelty I usually buy and sell bikes fairly regularly. Most bikes stick around for 1-3 riding seasons on average. However, one bike I can never bring myself to sell. It is the first motorcycle I ever bought (back in 2014) a 1976 Honda CJ360T I’ve named Josie.</p>
</p>
As reliable as these old hondas are they are still 50 year old bikes and tend to require a fair bit of maintenance and tinkering, but that is half the fun. This post is my experience using tools in my home office to remake an out of production rubber part for my bike.</p>

The Part</h3>
On each side of the CJ is an airbox that holds the air filter and feeds air into the carbs. They are held onto the bike with 3 bolts (Part #20). To keep vibrations down there is a rubber bushing that isolates the airbox from the frame (part #16). Finally, there is a metal washer/collar piece that is the interface between the bolt and the bushing. Refer to the part diagram below for a visual representation.</p>

Diagram from Partzilla</a></figcaption> </figure>
The rubber bushings on my motorcycle are all either missing or in various states of disintegration. The Honda part number of the bushing is `31403-323-000</code> and while it is no longer manufactured, some suppliers do have new-old-stock examples available.</p>`
Picture of the original part courtesy of PartsWarehouse</a></figcaption> </figure> They seem to go for $4 each with shipping on top of that. I need six of them which would come to $24 plus shipping (</del>$8 when I looked). This got me thinking about trying to produce these parts myself. These are not the only degrading rubber parts on a bike as old as this and some of those parts aren’t available anymore at all. So a success here would allow me to make quite a few parts for this bike.</p> The Plan</h3> My plan was to use 2 part silicone and cast it in a 3d printed mold. This type of silicone is most often used for making molds however a recent video on the FormLabs youtube channel</a> got me thinking about using it with a 3d printed mold to produce a final part. I ordered a cheap kit on Amazon with seemingly decent reviews, and started 3D modeling while awaiting its arrival.</p> Mold V1</h2> Design</h3> I went to Fusion 360 and started by modeling the bushing itself. From there I modeled a cube over the part and subtracted out the bushing from the middle. This left me with a solid cube with a negative space in the middle in the shape of the part I wanted. I split this cube into 3 parts, one cylinder for the middle and two halves that hold the cylinder in place. The two halves got a set of four holes so they could be bolted together with M3 hardware. For each half I added a hole into the molding chamber, one for silicone to be injected into and the other for air to escape from. To that end, I placed one hole on the bottom and one on the top.</p> </p> </p> </p> My plan was to use a large syringe to inject the silicone into the mold. I sourced a kit off amazon that had large syringes and a length of tubing. I sized the holes on the outside of the mold to accept the tubing.</p> Mold V1 after printing in Gray PLA on my Prusa MK2</figcaption> </figure> Casting</h3> Upon receiving my silicone, I jumped right into trying my mold out. After filling the mold the part needed to cure overnight. However, I quickly realized I needed some way to seal off the fill hole of the mold. I ended up wrapping the whole mold in strips of duct tape to try and seal off both holes, but the duct tape did not stick well to the silicone covered mold and the seal was anything but liquid tight. Additionally, while the 2 part silicone claims a 6 hour cure time I found that it was still very liquidy even after 6 hours (based on the leftovers in the mixing cup). Looking at reviews for the product this seems to be a common complaint. Thankfully, since this is platinum-cure silicone the cure can be sped up with heat. I put the mold and mixing cup of leftovers in a warm spot and by the next morning they were set up.</p> Results</h3> Finally it was time to de-mold the parts and the results were less than ideal. However, that was to be expected for a first shot, and the lessons I learned allowed me to quickly iterate my next attempt.</p> </p> A large amount of air ended up being trapped in the mold and because of that ~1/3 of the part was just missing. I think this had two causes; first, the poor seal of the duct tape covering on the fill port let some silicone seep out and air took its place. Second, bubbles in the silicone mix rose to the surface becoming one large air pocket.</p> The first issue can be addressed with a better designed mold. For the second, I do not have access to a vacuum chamber which is the proper way to deal with bubbles in the mix. However the mix I am using advertises its ability to degas itself after pouring. This is likely accomplished with its relatively low viscosity and long cure time.</p> Mold V2</h2> Design</h3> Armed with the lessons learned from my first mold design and my first casting experience I tried again. This time I decided to make the mold with an open top. This would allow bubbles in the mix to rise to the surface without getting trapped. Pouring the mix in from an open top would also mean there is no fill port to seal.</p> The new design started out just as before with a cube that had the model of the busing itself subtracted from the middle. Next I cut out the top of the mold with a circle just barely wider than the bushing itself (this was to create a lip I could use to gauge the height during the pour). Next, I split the cube with very shallow cone shape to create a separate top piece. I split this top in half lengthwise and added bolt holes. This gave me 3 pieces; a bottom with the pug for the central hole and two top halves that could be pulled away from the side. By pulling away from the side the mold could cast the undercut of the bushing shape and still be pulled away after curing.</p> Finally, I decided to try for 2 parts in one go this time since the smallest batch of mix you can feasibly make is quite a bit more than what is needed for a single bushing, and any leftovers are effectively just a waste. To do that I just copied all three pieces, shifted the copies over and combined them to the original bodies.</p> </p> </p> </p> Casting</h3> This time casting went much smoother. I let the silicone sit for about 30 minutes after mixing so that most of the bubbles could dissipate before I tried pouring it. I then poured the mix into the mold, pouring slowly, filling the bottom completely before letting any silicone get into the top. I did this to avoid trapping air under the shelf in the mold. The seal between the mold pieces isn’t perfect but the mix is viscous enough to not leak out. Finally, I put the poured molds on to my 3d printer’s build plate with the heat set to 45°C and placed a cardboard box over them.</p> Results</h3> After the pieces were fully cured I was able to remove them from the mold without too much difficulty and the results were much better. Below is a picture of the two pieces after I trimmed the flashing off of them.</p> </p> There are still quite a few bubbles in the parts, however they seem solid enough to be usable. For my next attempt I will probably let the silicone sit even longer before pouring it and see if that helps.</p> Wrap Up</h2> I have yet to try these parts on my bike. I need to cast 4 more and then I will try them out. I have some concerns about their durability as they quite a bit softer than I imagine the OEM parts are but only time will tell. Either way I am quite impressed with how easy this process is and already have at least one other part I want to try remaking. I am also planning on looking into pigments for 2 part silicone so I can make the parts black and get a more OEM look. I will try to post an update on how these parts work in the field and any other parts I end up making with the same process.</p>

Hosted UI</h2>
If you end up building one of these you can access the UI Page here on my website by scanning the QR code below of going to the URL https://scatena.co/pressure_sense</a></p>
</p>

Mold V2</h2>

Jason Scatena

Mechanical Autopen Simulation

Bluetooth Manometer for Carburetor Synchronization

Internal Mic Fix for Lenovo T14s Gen 4 w/ AMD on Ubuntu 24.04

Molded Silicone Parts Follow Up

Molded Silicone Parts for a Vintage Honda CJ360T

Jason Scatena

Mechanical Autopen Simulation

Computer Vision Detour</h3> To simulate the matrix, I needed a tool that could be run on the source image and create an output file with the curve information required to run the simulation and visualization. There ended up being 3 main concerns to deal with:</p>

Results</h2> </p>

Bluetooth Manometer for Carburetor Synchronization

Hosted UI</h2> If you end up building one of these you can access the UI Page here on my website by scanning the QR code below of going to the URL https://scatena.co/pressure_sense</a></p> </p>

Internal Mic Fix for Lenovo T14s Gen 4 w/ AMD on Ubuntu 24.04

The Issue</h2> In the Ubuntu setting I was unable to find any recording devices. However, running the command: lspci -knn | grep Audio</code> I was able to see three devices. Notably I saw:</p>

Molded Silicone Parts Follow Up

Molded Silicone Parts for a Vintage Honda CJ360T

Mold V2</h2>

Hosted UI</h2>
If you end up building one of these you can access the UI Page here on my website by scanning the QR code below of going to the URL https://scatena.co/pressure_sense</a></p>
</p>