When I started experimenting with AI vibe coding, I wouldn't have predicted that it would actually be useful. While I am still unsure about trusting whatever is returned by Claude for professional code, it has earned its spot for personal projects.

In the last couple of months, I have been able to knock off several items on my running tally of software projects. Most of the items are silly but are still worth it because the math has changed on how good they need to be. When the investment to write them goes down significantly (as with AI vibe coding), then building them trends towards zero.

Knocked off my list are 3 iOS apps, 2 native macOS apps, 2 web apps, and a Chrome extension.

iOS apps are something I have never put in enough effort to really learn. Every couple of years, I dabble at trying to make something trivial, but the amount of effort is always too high for the expected return. One great benefit to coding in Swift is that it is a compiled language, so I can instruct my AI coding assistant to keep chasing compiler errors and get back to me when it builds cleanly. This has made it a breeze to build with.

So far, I've built for iOS:

• A personal CRM to get better at remembering to talk with friends. Turns out there is no practical way to access call records or messages, so this died quickly. The fast iteration cycle let me try several alternative methods, but ultimately, it didn't work.
• A viewer for construction equipment for my toddler. Who doesn't love big trucks?
• An uncertainty calculator for exploring different outcomes when you don't have all the numbers to calculate. This may actually be released to the App Store.

For macOS, I have always just written Python. Trying to use more of the considerable system resources is always a challenge and usually not worth dealing with. I wanted to make use of all the cores/GPUs on my M4 Pro.

• A file system watcher to make sure my git repos are in sync. One thing I hate is having dangling changes or unchecked-in files on a local machine. Having a menu bar icon and a native way to watch FSevents is the best solution here. This probably won't get released, but it is incredibly helpful to run myself.
• A Tableau alternative for EDA (exploratory data analysis). This was much harder than I thought and falls into the deceptive category of problems that are easy to get started with but exponentially harder when dealing with real data. The goal was to make a tool to help make sense of a single project and instead got pulled into solving an even harder problem in a generic way. This got abandoned because it wasn't viable. I might have to try again if I get a better idea.

Web apps are my bread and butter. I don't need technical help writing them, but it is nice to have an AI assistant generate all the boilerplate when setting up a project.

• A web app to aggregate the screenshots folder across multiple machines. It runs on my Synology NAS, so it has access to all my data and is also available to all machines. I have a decade of screenshots and screen recordings that are nice to refer back to. It is a nicer way to access older demos without doing the YouTube dance.
• Personal bookmark keeper with a content recommender. This started so I could explore how to build a recommender. Using examples datasets from Kaggle is a struggle because I don’t have a sense for the quality of recommendations. The explore vs exploit tradeoff is easier to see when I understand the domain better. It has now expanded to pull in my RSS feeds. This will probably be released at some point because it is great to have. It only makes sense to be as self-hosted because I don't trust companies to manage my data responsibly.

The Chrome extension to remove shorts was written about here. . I don’t think it has enough appeal to deal with the Google review process. Extensions can access a ton of data and I don’t really want to have that liability.

In conclusion, at $20/month, it makes sense to have an AI assistant write a lot of these projects. For the low risk of errors in them, vibe coding is acceptable for making something good enough.

I don't know why I hate YouTube Shorts so much, but they do. I find it a very user hostile feature that even with a YouTube Premium account there is no way to disable it.

I get that that they can't resist the siren song of enshittification but that doesn't mean I have to see it.

Sure I could use one of the existing extensions but I don't like the idea of giving that much access to an unknown 3rd-party. So the natural answer is to write a simple version. Here is the first version of it that you are free you use, enjoy, change, etc.

You can find it here: https://github.com/mcotton/remove_yt_shorts

As an aside, this is the type of project that is perfect for vibe coding. It is low stakes and I don't really care about code quality for it. YouTube will continue to changing their page layout which means it is easier to just ask Cursor to investigate the selectors needed to hide the elements. Even with writing this blog post I am only 1 hour into the project.

I've been dealing with an annoyance for years. Having multiple computers means that you now need to coordinate where files are. This is a known issue for larger, important files. Where it is less obvious, are in the smaller transient files like screenshots.

I take a lot of screenshots and screen recordings. It is a cheap way to keep a snapshot of a moment in time. Got a new feature working? Take a screenshot. About to make some big changes? Take a screenshot. Want to share it with someone? Record a one minute recording.

All of these screenshots and recordings got into ~/Documents/screenrecordings and this is where the trouble begins. The common case is to backup the individual computers to a central location. The harder case is to synchronize back down to the individual machines. Without the synchronizing to the individual machines you can never know if you don't have the screenshot because you didn't take it, or if you just don't have it on this machine.

Introducing Listing which is a quick, lightweight web app that acts as a file browser for the centralized folder. It runs in docker and has a docker-compose file specifically for the Synology.

This isn't a big thing, but it removes some minor annoyances and hopefully can help someone else.

I am winding down the first phase of a customer pilot and wanted to celebrate a little before getting wrapped up in phase two.

This project needs some AI which is really data science, computer vision, and machine learning. There are several data sources that need to be integrated into a dashboard and data explorer. The interesting part is that this is a great use case for my prior classifier work.

The idea behind my classifier project is instead of having a single large model, with huge datasets, what about small models for each camera? A huge training dataset is needed for the model to generalize and prevent overfitting.

A security camera doesn't change often (if at all) and so it has less of a need to generalize. By embracing overfitting you can get great results with very modest training data. In fact, when training a binary classifier, you can get great results starting with tens of verified images in both classes.

If you iterate, correct misclassifications, and repeat this process, after several iterations you will have a semi-supervised classifier.

What are the downsides of this approach? By embracing overfitting, you run into the potential of misclassifications the first time your model encounters something new. It is unknown what will happen when it first encounters rain or bugs. This can be mitigated by continuing the semi-supervised learning approach. Anomalies and outliers need to be reviewed.

Each year we go and race in the Day in the Dirt down south race. It is a terrific event put on by RedBull. This year I decided to make commemorative trophies for my brothers and nephew. I'm happy with how they turned out and for Send Cut Send for doing the machine work.

It started as a quick sketch in onShape. I exported the two files as DFXs and uploaded to SendCutSend.

I order one frame in 0.250" aluminum, three in 0.250" steel powder-coated white, and one in 0.250" steel powder-coated red. This is how each of them arrived from SendCutSend. It only took some RedBull cans in various flavors and some epoxy to get them all assembled.

Here is how the prototype in aluminum turned out. I think it looks good and was a fun excuse to expriment for other subtractive maching processes.

First, the story

Once upon a time, in a pond on the outskirts of town, there lived a duck named Julio. Julio was a bit of a goofball, but he was also a very curious duck.

One day, while Julio was swimming through the pond, he overheard a group of swans talking about something called "cognitive biases." Julio had no idea what that was, so he asked the swans to explain it to him.

"Cognitive biases are ways that our brains can trick us into thinking things that aren't necessarily true," one of the swans explained. "For example, if we see something that confirms our beliefs, we're more likely to believe it, even if it's not true."

Julio was like, "Woah, that's messed up. My brain is tricking me all the time!" He knew he had to learn more about cognitive biases.

So Julio spent the next few days reading books, talking to other animals, and asking questions. He learned about all kinds of biases, like the confirmation bias, the anchoring bias, and the framing effect.

One day, Julio was hanging out with his best friend, a flamingo named Frank. Frank was trying to decide which type of worms to eat for lunch.

Julio noticed that Frank was only looking at worms that were pink, because he thought pink worms tasted the best. But Julio knew that Frank's brain was playing a trick on him. That's because Frank had already decided that pink worms tasted the best, so he was only looking for evidence that confirmed his belief. That's the confirmation bias in action.

Julio told Frank about the confirmation bias and how it could lead to bad decisions. Frank was like, "Woah, that's messed up. I need to be more careful about how I make decisions."

Julio was happy to see that his friend was learning and growing. He knew that understanding cognitive biases was an important part of being a wise and thoughtful animal.

As Julio swam through the pond, he couldn't help but laugh at the swans. "Those guys are always so stuck up," he thought. "They think they're better than everyone else just because they're white and have long necks. What a bunch of jerks."

And here is how I generated it

This is powered by ChatGPT and Dall-e from OpenAI. It all start with a text prompt on ChatGPT

Write a story for children, in the style of Mitch Hedburg, about a duck, name "Julio", that learns about cognitive biases, include an example of each cognitive bias, include text prompts for Dall-e illustrations. Include links to Dall-e illustrations. Include mildly rude jokes about swans.

ChatGPT returned this for the Dall-e prompts, I had to fill in the blanks a little bit

Text prompts for Dall-e illustrations:

1. Julio swimming through the pond, with a curious expression on his face.
2. A group of swans gathered around a lily pad, deep in conversation.
3. Julio reading a book about

So what do the suggested images look like?

I have been a huge fan of 3D printing and additive manufacturing. It is great for functional items like brackets, mounts, and replacement parts to fix a vacuum. The only problem, is that most people don't actual want stuff made out of plastic.

Switching over to subtractive manufacturing, and step back 50 years in time, brings us to CNC. CNC is an elegant technology in the sense that it is fairly simple, robust, proven to work, and incredibly flexible. The biggest problems with it are having to learn all the related woodworking sills and that most people don't want coasters any more than they want plastic widgets.

So why spend the time and money to learn it?

First, it is a robot cutting things that I design in a digital space. All of the annoyances of measuring, sawing, measuring again, goes away when it can be designed digitally first.

Second, it is a stepping stone to doing parts fabrication. Graduating up to machining metal is the real goal. The skills I'm learning now will directly translate up the to larger machines. Doing carbon fiber composite work needs molds and some molds are easier to CNC than to 3D print.

Third, working with wood is surprisingly satisfying. There is a whole community of people who woodwork and even casual acquaintances would rather talk about a nice slab of maple instead of hearing about the new LWPLA filament.

Looking at the image at the top of this post, the object itself is more interesting than the same shapes 3D printed. It just seems like more effort and skill were required. It did require more effort and skill and that is the whole point.

Skipping the hype around Docker, Kubernetes, and containers in general; I wanted to talk through how I use them and where they have been very helpful. The three use cases I want to highlight are local development, API examples, and deployment.

Local development with Python is a headache of dependencies and conflicting Python versions. You can use virtual environments or Docker containers, but docker containers have several other advantages when deploying or sharing. If you start from a fresh python image you will have to explicitly declare your module dependencies which becomes great documentation.

Providing API examples inside of a Docker container is another great use. When bringing people up to speed with a new API it is vital to get them early success and it is important to quickly get past the drudgery of getting an example to run. If you keep the docker container as a very light wrapper, there is very little confusion versus delivering them just the source directly.

Example API scripts for the Eagle Eye Networks Cloud VMS are available here on github.

I prefer to use Docker-compose for deployments in production. If there is a chance that I will deploy it at some point, than I start with docker-compose even when just running it locally. I find that the full DevOps-level tools are too much overhead for not enough gain. I feel the same about container registries. I haven't found the advantage to publishing the complete example back to a registry.

I run nginx locally on each instance for SSL termination and then use "proxy_pass" to send the traffic to the correct Docker container. This works great locally also because you can access those same entry points on localhost. The basic structure is shown below:

server {

    root /var/www/html;

    client_max_body_size 100M;

    server_name xxxx.een.cloud;

    location / {
        proxy_pass http://localhost:3000;
    }

}

Running it locally you can have all requests go to the single container running on port 3000 as shown above. When you are in production, you can split up your traffic by URL requests to go to additional copies of the same container. In this case, I want to work on my main webapp, ingestion API, and notification system to work independently. I can do this by modifying my nginx config and docker-compose.yaml as shown below.

server {

    root /var/www/html;

    client_max_body_size 100M;

    server_name xxxx.een.cloud;

    location /api/ingest {
        proxy_pass http://localhost:3001;
    }

    location /api/notification {
        proxy_pass http://localhost:3002;
    }

    location / {
        proxy_pass http://localhost:3000;
    }

}

version: '3'
  
services:
  app:
    build: .
    volumes:
      - .:/app
    ports:
      - "3002:3000"
    restart: always
    tty: true

  ingestor:
    build: .
    volumes:
      - .:/app
    ports:
      - "3001:3000"
    restart: always
    tty: true

  notification:
    build: .
    volumes:
      - .:/app
    ports:
      - "3002:3000"
    restart: always
    tty: true

This isn't the complete answer for how everyone should use Docker everywhere, it is just want I've found helpful. Hopefully it is helpful to others.

It is cheap and easy to build a machine with 8/16 cores and 32GB of RAM. It is more complicated to make Python use all those resources. This blog post will go through strategies to use all the CPU, RAM, and the speed of your storage device.

I am using the AMD 3700x from my previous post. It has 8 cores and 16 threads. For this post I will be treating it each thread as a core because that is how Ubuntu displays it in System Monitor.

Looping through a of directory of 4 million images and doing inference on them one by one is slow. Most of time is waiting on system IO. Loading the image from disk into RAM is slow. Transforming the image once it is in RAM is very fast and making an inference with the GPU is also fast. In that case Python will only be using 1/16th of the available total processing power and only that single images will be stored in RAM. Using a SSD or NVME device instead of a traditional hard drive does speed it up, but not enough.

Loading images into RAM is great but you will run out at some point so it is best to lazy load them. In this case I wrote a quick generator function that takes an argument of the batch size it should load.

Dealing with a batch of images is better than loading them individually but they still need to be pre-processed by the CPU and placed in a queue. This is slow when the CPU is only apple to use 1/16th of its abilities.

Using the included multiprocessing package you can easily create a bunch of processes and use a queue to shuffle data between them. It also includes the ability to create a pool of processes to make it even more straight forward.

In my own testing, my HDD was still slowing down the system because it wasn't able to keep all of the CPU processes busy. I was only able to utilize ~75% of my CPU when loading from a 7200RPM HDD. For testing purposes I loaded up a batch on my NVME drive and it easily exceed the CPUs ability to process them. Only having a single NVME drive I will need to wait for prices to come down before I can convert all of my storage over to ultra-fast flash.

Using the above code you can easily max out your RAM and CPU. Doing this for batches of images means that there is always a supply of images in RAM for the GPU to consume. It also means that going through those 4 million images won't take longer than needed. Next challenge is to speed up GPU inference.

Machine learning has changed how I approach new programming tasks. I am going to be working through an example of detecting vehicles in a specific parking space. Traditionally this would be done by looking for motion inside of a specific ROI (region of interest). In this blog post I will be talking through achieving better results by focusing on dataset curation and machine learning.

The traditional way to detect a vehicle would be to try and catch it at a transition point, entering or exiting the frame. You could compare the video frames to see what has changed and if an object has entered or exiting the ROI. You could set up an expect minimum and maximum object size to improve accuracy and avoid false positive results..

Another way would be to look for known landmarks at known locations such as lines separating spaces or disabled vehicle symbol. This would involve some adjustment for scaling and skew but it could have a startup routine that was able to adapt to reasonable values for both. It wouldn't be detecting a vehicle, but there is a high probability that if the features are not visible then the spot is occupied. This should work at night by choosing features that have some degree of visibility at night or through rain.

A third technique could be similar to the previous example expect looking at the dominant color instead of specific features. This would not work at night when the camera switches to IR illumination and an IR filter. This sounds like a lazy option, but in specific circumstances it could perform quite well for looking for brown UPS vehicles that only make deliveries during daylight hours

In the traditional methods, you would look at the image with OpenCV and then operate on pixel data. Some assumptions would need to made on how individual pixels should be group together to form an object. The ROI would need to be defined. Care would also need to applied when handing how individual pixels create a specific feature, or are composed a color between our threshold values. All of this has been done before.

The approach I am advocating for is to not program anything about the image at all. Each frame would be looked at and the model would make a prediction regardless of motion, features, or dominant colors. Such a general approach could easily be marketed as AI. I prefer to call it machine learning to clarify that the machine is doing the work and that it's only intelligence is in its ability to compare to known images. Think of it as a graph with a line that separates it into categories. The algorithm's only job is to predict where to place each image on the correct side of the line.

As more example images are added to the model, the line separating the groups of images changes to accomodate the complete dataset. If you are not careful to keep the two groups balanced than biases in the datasets can appear. Think of this as a simple programming bug.

The first image is considered in the parking space, the second image is not. In both cases, the vehicle is blocking the parking space. The difference is arbitrary, but these special cases can be handled by including examples in the correct dataset. Think of what goes into the datasets as domain specific rules.

It is easy to build a dataset based on images evenly spaced through the day. Doing this should give good coverage of all lighting conditions. This method also trains the model with a potential hidden bias. Because there are more cars parking during the day, it is more likely the model will learn that night time always means there is no car. The image on left shows a person walking and gets a different result than a nearly identical image during the day. The image on the right is included to balance the dataset to account for this bias.

The line separating the two groups is getting more complicated and would be very difficult to program using traditional methods.

These final examples show that the model must be able to account for unexpected events such as heavy rain or a car parked sideways. Handling unexpected data like this using the traditional methods would require significant rework.

In building this model, the programming stayed the same. It is all about the data and the curation of the datasets. The future of programming will be based on curating datasets and less about hand coding rules. I'm very excited for this change in programming and the new applications that can be made through it. We will need our tooling to catch up and I am excited to be working on it.

Questions, comments, concerns? Feel free to reach out to me at mcotton at mcottondesign.com

The cloud is a great way to get started with AI/ML. The going rate is around $4/hr for a GPU instance. It might be all that you need, but if you need to maximize your personal hardware buget this is my guide to buiding workstation.

Price

I spent ~$1000 and you could get it even lower with the current deals. I decided on a AMD Ryzen 3700X, Nvidia 2060 Super, 32GB of RAM, and an NVME drive. I could connect all of this into a B450m motherboard so I didn't see any reason to spend more. I also include another SSD for Windows 10 and two HDDs for storing datasets.

CPU

The Ryzen 3700x is far more than I need and most of the time several cores are idle. Because most of my tooling is in python it is a real struggle to make use of the resources. The multiprocess library is great and it isn't too complicated to make a Pool of CPU cores for the program to use.

GPU

The 2060 Super is in a very odd position in the current lineup. It is the cheapest model with 8GB of VRAM, to get 3GB more you have to 3x the price. It makes more sense to jump up to 24GB for 6x the price. Training CNN models on image data requires significant amounts of VRAM. You could spend more but without more VRAM the money would only provide marginal improvements due to the number of CUDA cores.

RAM

System memory is very similar to CPU cores, you will rarely max it out but it never hurts to have more available. I typically am using 20GB which means that I could reduce my batch size and be done with 16GB or I could just spend the extra $80 to get 32GB.

STORAGE

I'm using a 1TB NVME drive. It is very fast and tucks away nicely on the motherboard. It is overkill for my needs but it is nice to have the space to store millions of images and not wait on traditional HDDs.

Speaking of HDDs. I have two, a 1TB and a 2TB. The 1TB is just for datasets and so that I can keep space available on the faster drive. The second drive is for automatic backsups of the main disk and datasets. Backups are also doing to two different NASs but that is another blog post.

It would be a shame to have this machine and not play games on it. I'm using an inexpensive 1TB SSD for Windows 10. I don't like dual booting off of a single drive. I prefer using the bios boot selector to switch between OSes.

COOLING

The case I'm using has four 120mm fans. I added a giant Nactua heat sink with two additional 140mm fans. Adjust the motherboard fan curves so that everything is nice and quiet. I believe in having a lot of airflow and I haven't had any temperature problems (inside the case, the damn thing very noticably heats up my office).

UPGRADES

I'm currently happy with the way it is. The obvious upgrade will be for more GPU VRAM. I decided against water-cooling but if I get to a place with multiple GPUs that appears to be the best solution.

This is a quick example on how to recognize QR codes in video streams and extract the metadata. Read the full write-up over at https://een.cloud.

This is a quick video on how to blink a light when an Eagle Eye camera detects motion. There are plenty of ways to do this, and this is just the way I decided based on what I had available on my desk.

I used the following parts:

• Raspberry Pi 2B
• Arduino Uno
• one red LED
• 3D printed Eagle Eye logo
• EE-blinker that runs on the Raspberry Pi
• blink_sketch that runs on the Arduino

The Arduino was configured with the red LED on pin 13. The sketch looks for serial communication (over the USB port from the Raspberry Pi) of an ASCII character. If the character is '1' then it turns on, if it is '2' then it turns off.

As an aside, why didn't I make it 1 or 0? Because originally I was planning on having multiple states and the states did a sequence of things.

The Raspberry Pi is just running a Node.js program and it could be replace by any other computer. The program itself could also be replaced with any other language to subscribe to the Eagle Eye API.

You can see the code for the entire project here. Make sure to read the README file and follow the instructions to create an appropriate `config.js` file. It also contains the correct serial port for connecting to the Arduino so make sure you have that correct.

The above example is listening for motion events. Specifically it is listening for ROMS and ROME events (ROI motion start, ROI motion end). It could easily be adapted to use other events.

If you are monitoring for status changes (camera online/offline, recording on/off, streaming on/off) you most likely want to listen the status bitmask directly. The events are delayed because we go through our own internal heuristics to filter out false positives.

You can find an example of subscribing to the poll stream, getting the status bitmask, and parsing it in this example project.

Introduction

I wanted to know if the garage door is left open. This is how I trained my computer to text me when it has been open for too long.

Detecting if a garage door is open or closed is a surprisingly challenging problem. It is possible to do this with contact sensors or even with a motion (PIR) sensor, but that requires running additional wires and a controller.

My solution was to put a wireless IP camera to the ceiling. It connects to my home network using wifi and plugs into the same power outlet as the garage door opener.

Traditional video security analytics such as motion or line cross does not adequitely detect the door being left open. Both motion and line cross depend on something happening to trigger. This is different because I am looking the position of the garage door instead of movement.

I also looked at a traditional computer vision solution to detect sunlight coming in throuogh the open door. Ultimately trying to detect based on the 'brightness' of the scene would not work in this situation. The way the camera is position it would not be able to differential bewteen the light from the ceiling and sunlight coming in through the garage door. This also would not work at night or during a storm. A different tool is needed.

I wanted to experiment with creating a machine learning model. This is a great first project to get start with.

This is a present I made for a friend. It makes a handy travel stand because it is cheap to produce, light weight, and snaps together with magnets.

You can download the STL files on Thingiverse

This is a kit that I designed and printed. Each bomb bay requires a channel on your RC receiver. In this video I am using an 8-channel receiver but you easily disconnect your rudder and use that while you fly the plane with just bank and yank.

You'll be able to download the STL files.

At a previous startup, I had a great idea that we could improve our product with realtime updates from the server. At the time we had implemented long-polling in the client. After failing to get everyone on board I decided to take a weekend to create an example with socket.io and Node.js.

On Monday morning I was proud to show it off around the company. The CEO liked the idea and agreed with me. Yay! We got on a call with the CTO and I proudly explained my work.

He looked at it, then asked me to go into further detail about my choices. How did I implement this? Tell him about websockets? How can we roll it into productions? What does it look like at the HTTP layer? Well, I could sort of answer his questions. I had to confess that I really didn't know and I used socket.io and it uses websockets.

He probed a little more and then started to get impatient. I couldn't believe that he had become dismissive. And worse, the CEO followed his lead. I hadn't figured it out yet but I had made a critical mistake.

Would we now be implementing the socket.io abstraction? Is this what we want? What trade off does that mean for us? Is this the only way?

I was afraid to learn how the technology actually worked and I was asking the company to take in several major dependencies in order to implement this.

I stubbornly maintained my position. Why would we skip over an existing answer? I had proved it was working. They could see it working. Why were they not as excited as I am?

We were at an impasse and I still hadn't figured it out yet.

Fast-forward to now, and I am a first time CTO. I've spent a lot of time thinking about lessons learned from my friend, the CTO in the previous story. I do my best to consider why he made the choices he did.

I'm interviewing candidates and I'm starting to see a similar pattern with other developers. They gladly accept dependencies and baggage that limits options and forces a direction. They'll make a terrible choice to avoid doing something that makes then uncomfortable.

Coders are afraid of writing code.

A fiend of mine recently described how their company need to do multiple events from an external webhook. Not enough people understand a simple web server so it got pushed to DevOps to build using salt. Salt isn't a horrible answer but when he was telling me about this I could see a little of my inexperienced self in the story.

On another project I was dealing with an Indian iOS developer that needed to sort an array. They fought me when I filed a bug that the sort order wasn't correct. A week later they had it working but had implement it in the worst way possible. They had pulled in SQLite as a new dependency and were using that to sort the array. When I saw this I was furious. They were very cheap agency but after this they weren't worth keeping around at any price.

In conclusion, we ended up implementing websockets. I was the champion for this feature and after our CTO guided me in the right direction. We read through the spec for websockets, we came up with the right fallback strategy, and the product improved significantly. We didn't have to blindly add dependencies, we understood the technology we wanted tovuse, and we found the right fit for our needs.

I was asked this once at an informal interview and didn't have an immediate answer. I told a story about the time I was trying to debug RF signals and the ultimate cause was that there was another remote using the same hardware ID. It was extremely unprobabel, but it turned out to be the cause.

Since then I've had more time to reflect on how I would answer that question if I was asked again.

The weirdest bugs I've ever dealt with all have a common theme, distributed systems. In my last two startups we have used distributed architectures. This has caused somewhat simple tasks to become very difficult.

In the old days of a monolithic app server (Django, Rails, PHP) all the logs were in one place. As you scaled you would need to add a load balancer and then spread the requests over multiple app servers. Reasoning about the path an HTTP request took was still fairly simple. At this level a relational database is perfectly able to handle the load.

In a distributed system, HTTP is used not just for the inital request but also for communication between servers. This means you now have to deal with internal HTTP responses and they have different meansing from the external request/response that you're used to. For example, what to do when an internal call to the cache returns a 401, 403, or 502? Do the unauthorized/forbidden calls mean that the external's request is 401/403 or does this mean your app server is? What about 502 Gateway Timeout? You now have to deal with timing and latency between internal services.

Distributed computer and networking are a reality of life for most new projects. It is something we'll have to learn and deal with. In the meantime, it is also the source of my weirdest bugs.

Showing off our truck on a nice, sunny afternoon. Shot with a Phantom 3 Pro and an Inspire 1.

I am excited to announce that I have accepted a new position with HUVRDATA. After an eight week transition I am thrilled to get started. I wI'll be their CTO and will guide them into becoming the leader in drone based inspections.

It was surprisingly hard to leave Eagle Eye and specifically, let the team know what was next for me. I lost sleep over this decision but we were able to make it through. I loved working with them and it was filled with great memories.

I've been working with Bob and Ben at HUVR since the start of the year. Things really kicked into gear when they closed a funding round in August. Right now it is time to mash the accelerator and I would have regretted letting them go on without me. This is an area I am very passionate about, have the skills to do it, and am fully on-board to make this successful.

Improving takes time and patience and lots of near-misses. There is usually a narrow window in to capture the ideal scene and if you aren't in position and ready you will miss it. I continue to be 10 minutes late, out of battery or space on the SD card when I realize this is the shot I wanted.

In this case, I wanted to try the waters before going again the next day. The challenges of shooting downtime is that you attract too much attention to play around with expensive equipment as a crowd gathers.

Pushing the distance out a little further and keeping it low across the water. Experimenting with Final Cut Pro X to get the titles and transition.

I had a terrific time with @LSRiverboats getting this footage. This was captured in 4k@30fps but rendered to 720p for Vimeo. The initial shot was digitally zoomed in and the reset is right off the camera. Some minor exposure correction was done in post to make every match.

This was a lot of fun but extremely challenging to operate from a boat. There were two boats, fading daylight, bats and other groups to contend with. Thankfully everything worked out and I am happy with the results.

I've been trying for quite some time to get this shot. It was interesting because this was also my longest distance flight. I only went 1000ft' up the river away and it took a lot of squinting to keep it in sight.

This was taken out front of the Austin City Power Plant and is using the new manual camera controls of the Phantom 3 Pro.

It is hard to go back to past locations and try to get a fresh take. The challenge is to that you are no longer looking at it with fresh eyes. This is made worse if you liked the results from previous tries.

I am challenging myself to throw away my previous work. All locations (except the Capital) are fair game. I am also going to not get hung up on having done the shot before. I will also shamelessly steal good ideas for others. I am still in the beginner phase and am in no position to be selective.

This is video from my first flight with the new Phantom 3 Pro. I picked it up earlier today and ran it through its paces. So far, it is a significant improvement from the Phantom 2 Vision +. The iOS app is the major differentiator. All the other features are in support of the app. The remote is the same as the Inspire 1 and feels much more refined. Common features can be done using hard buttons and there is less time spent pawing at the app. I am a big fan of physical buttons and not having to look at what I'm pushing.

I have (foolishly) upgraded my iPad Retina mini to beta versions of iOS 9. The DJI app isn't playing nicely with the new HD video streaming. I can't hold this against them.