artificial intelligence Archives

May 21, 2026May 21, 2026

Vibe-coding is a Game Changer

Vibe-coding has become a game changer since I wrote about vibe-coding as a non-developer a couple of months ago. Since then, my coding activities have expanded exponentially so I figure it’s a good time to write about how my use of AI coding has evolved.

The Statistics.

Let me start by showing my activity records on GitHub and (local) Gitea since it will set the tone for the rest of the post. The short version: an explosion of activity from mid-March onwards.

Gitea:

Github:

With the help of AI I was able to set up a simple home network setup with two devices: one old X86 laptop and one Orange Pi 5. These two devices serve as a local backup storage and run a couple of helpful tools, including my local Gitea which I use to keep track of my code projects.

With the help of my brother I also set up a Tailscale net which means I can commit to my Gitea repositories from home, the office, or any coffee shop I happen to work at.

The Projects.

My coding projects fall into two categories: personal and professional. I can’t share every project I’ve worked on the past six months but here’s a good selection.

VoiceTask

I extensively use Asana to keep track of all my personal and professional tasks for the various companies and projects I’m involved with. It’s a great way to keep things organized when sitting behind the laptop but I struggle to add tasks if I’m on the road.

So, I vibe-coded an application that records my message, transcribes it, then sends it to my Asana inbox for that day. When I get back behind my laptop, I see the task, and can re-assign accordingly.

After transcription I can edit the task title

I’m at v29 already so there’s been quite a few bugs squashed over the past weeks. Early on I made the decision to create it as a progressive web app (PWA) which means the app now lives on my phone as an icon. The transcription is handled by the Orange Pi 5 with its 8-core Rockchip RK3588 and the service is exposed as a Tailscale funnel. And the whole thing was up and running in under 24 hours too!

Museum of Overclocking Records

I’ve been collecting information on the history of overclocking world records for years – first in my professional capacity at HWBOT, now in my hobby-capacity at SkatterBencher. While I was happy with the historical timelines I hosted on my blog (CPU, GPU, MEM), it always bothered me I couldn’t easily link assets and sources to the records. With the help of AI I was finally able to put together a custom interface for managing the information and publishing it on the internet.

Each record is stored in its own folder which contains a JSON file with record details and all related assets like pictures. Then, we deploy the website via GitHub pages. The deployment involves building the website assets, resizing the pictures ensuring ease of browsing, and calculating the timelines and a bunch of statistics.

I had never even heard of GitHub pages before AI suggested it’s a good way to host a website. And, now, I can’t imagine doing it any other way. What’s also helpful is the admin page I vibe-coded which provides a user interface to manage and add records.

BENCHLAB

BENCHLAB is a tool that provides a way to collect system telemetry for PC desktops. I’ve also been using it as a personal vibe-coding sandbox. There’s too many little things I’ve developed for BENCHLAB, so here’s a short summary:

PyCore: python library that handles COM serial for the BENCHLAB device.
PyTools: a number of small Python tools that facilitate interacting with the BENCHLAB in Windows, Linux, and with X86 and ARM devices. The tool list includes: restAPI server, a TUI, CSV logging, exporting to HWiNFO, communication with MQTT brokers, and displaying telemetry on the WigiDash and VU Dials.
Windows Service: expanded the existing Windows service with multi-device support, added support for an upcoming variant (shht!), and added the ability to program the BENCHLAB device.
ICue Widget: an iCue widget that displays BENCHLAB telemetry on the CORSAIR Xeneon Edge.

Microsoft Graph API

A project I can detail even less is the migration to Microsoft 365. The specifics aren’t important but what matters is that Microsoft provides a useful Graph API that enables interacting with its various applications.

With the help of some AI and vibe-coding effort, I was able to generate automatic reports and emails for internal Planners.

The Cost.

Probably the most important topic is cost of all this vibe-coding activity. Spoiler: the grand total comes out to about the price of a dinner out.

Yes, it’s true, since my last blog post I’ve started spending money on imaginary AI tokens. My total token expenditure so far is US$50 with OpenRouter and US$25 with Cline, though I still have about US$30 in credits left. Of the US$45 worth of tokens I spent, I embarrassingly admit about US$20 of that went to a single Claude Sonnet 4.6 “review this repo” request by leaving sub-agents enabled. That was an expensive mistake!

So, I’d say the token cost of all my vibe-coding projects was about US$25. That’s next to nothing compared to the value it generated for me. I often find myself repeating to folks it’s a crazy world we live in.

The Setup.

Before I sign off, let me quickly detail my current setup because it’s not at all what you’d expect.

My daily carry is an older Dell XPS laptop which I definitely need to swap out for a newer model sometime soon. I have switched to a real IDE, VS Code, which I now use for all my development efforts. The only plugin I use for vibe-coding is Cline and it has served me incredibly well.

Originally, I used the free models provided by Cline – my favorites being KAT-Coder-Pro V1 and Qwen3-coder – but after a while I found the service to disconnect too often. That’s when I signed up at OpenRouter. But lately I’ve found myself back at Cline and paying for KAT-Coder-Pro V2 which gives me satisfactory results.

The Workflow.

Now, to kick off a project I actually often turn to Claude.ai because it’s exceptional at providing no nonsense, multi-file initial drafts of a project, and because it’s free. As long as the project’s not too convoluted, you can use it for long enough to build something that works, then port it over to VS Code to further build it out with Cline.

Often, I’ll also go to Claude when there’s a nasty bug I can’t seem to get rid of. More often than not, Claude finds the solution.

I used to rely on ChatGPT a lot. I still try it once in a while but I find it to produce too much fluff and emojis. For example, I would ask it to find a bug and then it would give me a solution … followed by “but here’s a better way”. It’s a waste of tokens and a waste of my time.

The Summary.

There’s no other way to put it: vibe-coding is a game changer for me. I can get computers to work for me at near-zero cost.

That may be a bit of an exaggeration but it’s how I feel about this new world of AI vibe-coding. Of course, I’ve had computers work for me for many years since every software lets me do something but this feels different.

It’s probably because the time to a personalized solution is much shorter. Yes, there are probably software tools I could use to get my voice message ported into Asana but these would require configuration and might not do exactly what I want or work how I want. But with vibe-coding, I skip the configuration entirely and get something tailored to my workflow.

And, of course, there are tools out there that let you operate a digital museum. But they need me to adjust my workflow to their interface. With a little vibe-coding effort and a couple of bucks spent on tokens, I was able to make a system that’s catered to my very specific needs (and niche).

And, of course, I could read through the Graph API and figure out how to get a task added to the Planner. But it’s just so much faster to write:

review @/documentation\plannerApi.md. I want to build a tool using the same credentials that enables me to create a new task in a planner.

and wait a couple of minutes until my script is ready.

Truth be told: I’ve produced a lot of code but I still don’t consider myself a software developer. So, my experience as a non-developer may not reflect how AI and vibe-coding is changing the world of software development. To learn more about that I recommend reading through Jared’s blog posts.

But if you’re a non-developer sitting on a problem that’s been nagging at you for years, my advice is simple: just start. The barrier has never been lower and the value never been higher.

October 12, 2025October 28, 2025

“Vibe Coding” as a Non-Developer

Vibe coding lets non-coders write functional software tools, though knowing how code works helps be most effective.

Preface

Let’s talk about “vibe coding” – a term coined by Andrej Karpathy not even a year ago, and already buzzing through the tech world. At its core, vibe coding means using large language models to build software by describing what you want in natural language, rather than wrangling complex syntax.

It’s kind of wild: suddenly, regular folks can spin up surprisingly capable tools without a traditional coding background or know-how.

My own “vibe coding” journey began in 2024. While writing a Raspberry Pi 5 overclocking guide, I used ChatGPT, CoPilot, and Gemini to help build a basic telemetry tool. The good news? It worked. The bad news? My brother (a seasoned developer) called the code “flimsy” and said it could be refactored into something far more robust (which he ended up helping with – thanks!).

Three Vibe Coding Summer Projects

So, this past summer, I wanted to see just how doable it was to use these large language models for real coding tasks as a non-developer. I took on three projects:

Orange Pi 5 Max Telemetry Tool
Rockchip RK3588 OC Tool
BENCHLAB PyTools

Orange Pi 5 Max Telemetry Tool

This project was a spiritual sequel to my Raspberry Pi 5 experiment as I worked on an OrangePi 5 Max 16GB overclocking guide. The goal: collect telemetry from the Linux OS and SoC, then display it in a simple terminal UI. Nothing fancy—just a clean test of what vibe coding could do with plain English prompts.

The result was a single-file Python script that provides a full telemetry dashboard for Rockchip RK3588-based boards like the Orange Pi 5 Max. It tracks CPU, GPU, NPU, and DSU clock speeds, system voltages, temperatures, load averages, and SAR-ADC readings.

The interface uses curses to present the data in a tabbed layout, with optional CSV logging for long-term analysis. Clock frequencies are decoded directly from hardware registers via /dev/mem, translating PLL sources and divider settings into actual operating conditions.

Rockchip RK3588 OC Tool

The next project was a step up in complexity. Rather than reading the SoC telemetry, I wanted to use register access to adjust clock frequency parameters at runtime.

This tool, also a single-file Python script, uses curses for its UI. The challenge was ensuring safe register access, meaning writing the right bits to the right registers while following Rockchip’s register writing rules, and avoiding system crashes. For example, enabling writing bits to the lower bits by first writing to the upper bits.

The result is that we have a tool that allows us to change the clock sources, adjust Rockchip’s PVTPLL configuration, and helped me overclock the SoC to over 3450 MHz.

BENCHLAB PyTools

The BENCHLAB PyTools project takes things even further as I wanted to explore working on a multi-file, complex software stack. The goal was to turn some BENCHLAB Python sample code into a handful of tools that bridge BENCHLAB telemetry with a desired output format, whether you’re on Windows or Linux. Here’s what it ended up doing:

TUI: Sensor monitoring via a text interface
CSV Logging: Sensor logging for data analysis
Graph: Sensor visualization using dearpygui
MQTT: Message BENCHLAB data to an MQTT server
VU: Display sensor output on a Streacom VU1 dial

Each tool shares core logic for serial communication and sensor parsing. They support multiple BENCHLAB devices and (should) run seamlessly across platforms.

I’m using this as a platform to explore what vibe coding can help with so I’m actively working on adding more functionality. Furthermore, the BENCHLAB team is working on some new service offerings that rely on the vibe-coded tools. Stay tuned for more!

What I Learned from Vibe Coding

The biggest win? Accessibility. I’m not a software developer – not even close – but with the help of AI I could build working tools in under a week. That’s the magic of AI: it lowers the barrier to entry and lets non-coders fill gaps in their workflow or hobbies with custom software.

But let’s be real: the code isn’t perfect. It’s rough, not super scalable, probably insecure, and definitely buggy in places. Still, without vibe coding, I’d first have to spend months learning to code before even getting started on a basic “hello world” project. Now, I can get straight to making the tools I need.

As projects grow, complexity creeps in. You’ll need to understand your code to debug it. Language models sometimes rename functions or introduce quirks that trip you up. It doesn’t always remember relevant context or disregards the important parts. And once you’re juggling issues involving multiple files and dependencies, things can get annoying.

One of the biggest pain points? Guidance. AI models still need a lot of hand-holding for tougher tasks. It can be slow and frustrating but that also makes it surprisingly effective as way to learn.

Final Thoughts

Vibe coding has opened doors for me that I thought I’d never walk through and it doesn’t cost a dime if you’re using free services. That’s huge. Of course, it’s not a silver bullet that will do all the hard work for you. You’ve got to know what you want and be ready to get your hands dirty. The clearer your goals, the better your results.

Don’t get hung up on fancy “prompt engineering” either. Just figure out what you want, ask for it, and tweak as you go. If you’re planning to launch a business around a vibe-coded project, though, definitely have a pro review your code for security, scalability, and stability.

Otherwise, have fun! It’s a powerful new way to build things, even if you don’t call yourself a coder or developer.

March 12, 2025March 12, 2025

Solutioneering the Transformation from Digital to AI Facilities

Technology-first solutioneering serves facility operators navigating the transformation of Digital into AI facilities.

Facility Management Transformation

Facility management and maintenance have undergone significant transformations over the past decades, from labor-intensive manual work to automated smart buildings, driven by technological advancements and changing operational needs.

AI and machine learning technologies are poised to accelerate this transformation, putting further pressure on facility operators and facility solution providers to drastically expand their services and deepen their skill set in order to keep physical and digital assets up to date.

The Olden Days

Before the rise of modern technology, facility maintenance was primarily manual, physical labor with rudimentary tools. Devices and equipment were repaired or replaced as they broke down and inventories were kept track of with pen and paper (or on a whiteboard, as below).

analog solutioneer — Superintendent of Buildings and Grounds Harold M. Wadsworth (left) and custodian foreman Dean Gardner of Utah State College select the crew staff for training in 1955. (facilitiesnet.com)

As industrialization and urbanization accelerated, the introduction of mechanical and electrical systems revolutionized facility management. HVAC systems, automated elevators, centralized electrical grids, and early security systems were integrated into buildings, requiring more specialized skills to maintain these systems. During this period, computerized maintenance management systems (CMMS) began emerging, allowing facilities to track and schedule maintenance more efficiently.

This evolution demanded an initial shift in services offered by facility solutions providers, moving away from providing hands for manual labor to bringing the expertise to operate electrical systems and computers.

The Digital Revolution

With the rise of advanced and connected computing, facility management saw another major shift towards digital solutions. The integration of enterprise resource planning (ERP) systems, sophisticated CMMS platforms, and building automation systems (BAS) now allowed facility managers to monitor and control multiple building systems remotely.

smart building — Advantech WISE-PaaS iBuilding Solution (advantech.com)

Energy efficiency became a growing concern, leading to the introduction of energy management systems that optimize heating, cooling, and lighting based on usage patterns. This decision-making forced facility operators to increasingly rely on digital information and data systems, which facility solution providers would develop and maintain for them.

The skills required to manage facilities dramatically shifted from craftsmen to skilled professionals familiar with digital technology stacks. The fundamental shift to technology-first facilities solutions and maintenance laid the foundation for an accelerated path towards smart and connected buildings.

The Internet of Things (IoT) has transformed facility maintenance, making real-time monitoring and predictive analytics an industry standard. IoT sensors collect data on equipment performance, environmental conditions, and occupancy patterns, which in turns enables us to develop maintenance strategies to anticipate equipment failures.

AI Transformation

Facilities are poised to be transformed by the impact of artificial intelligence in the coming decade. Aside from cloud computing, digital twins, and smart energy further enhancing energy efficiency, reducing operational costs and minimizing downtime, on-premise “AI workers” will take over more and more facility floor space as they assist companies to grow and expand more cost-effectively.

ai solutioneering — An AI generated image of a modern office where a human and a humanoid robot are interacting (dall-e)

These “AI workers” aren’t just robots sitting alongside the production assembly line, but are also security drones monitoring the premise, call center agents solving customer problems, internal sales assistants providing training to sales staff, quality assurance teams ensuring faulty products are caught before shipping out, and so on. This is challenging facility solution providers once again as it requires a new layer of skills to be offered to facility operators.

These new skills include not only assembly and maintenance of AI systems, but also providing AI solutions that ensure the “AI workers” continuously learn from their human counterparts working at the company. To remain competitive, it’s no longer good enough to rely on an off-the-shelf software solution or SaaS; proprietary business processes that provide a competitive edge in the market must be encoded within AI models and reinforced with your team’s best practices, innovative ideas, and lessons learned.

That’s the particular challenge MTS Solutions is tackling on behalf of our customers as we help our customers navigate the exciting transformation of Digital into AI facilities.

February 8, 2025February 5, 2025

Is AI Labor-Serving or Labor-Saving?

Is AI supposed to be labor-serving or labor-saving?” is one of the most often asked questions when talking about AI transformation. Time and time again, we (at MTS Solutons) emphasize the same point:

AI is meant to be labor-serving, not labor-saving.

Henry Ford – founder of the 121-year old Ford Motor company – addressed a similar concern a century ago. As the mechanized assembly line revolutionized industry, many feared machines would replace workers entirely. In 1929, Ford put it plainly [1]: “For unless machinery is labour-serving, it has no excuse for being.” We can adopt that sentiment and say:

“For unless artificial intelligence is labor-serving, it has no excuse for being”

We see the same fears today with AI. Some predict that artificial intelligence will inevitably take over many human jobs, leaving workers with no place in the economy. But history tells a different story.

When stronger, faster, higher-powered machines emerged in the early 20th century, it didn’t eliminate work — it transformed it. Businesses became more productive, industries expanded, and new jobs emerged that were previously unimaginable. Furthermore, the new jobs required more human intelligence, not less.

Workers at River Rouge, Courtesy: The Henry Ford Museum

Ford also noted that the power generated to run the machines by itself, is meaningless. You need people to extract the benefits from the abundance of power through the use of machines.

The same applies to AI today. The raw computing power of modern AI chips is useless unless it’s used to develop algorithms that enhance human productivity, helping people perform their jobs better, not eliminating them.

That said, of course some tasks and jobs will be fully automated by artificial intelligence — but only where automation makes sense. A good rule of thumb is this: if a task requires no judgment, no creativity, and no decision-making, then it’s better handled by AI. There’s no point in having a person spend time on repetitive, mindless tasks when they could be focusing on higher-value work — work that requires out of the box thinking, intuition, problem-solving, or relationship-building.

“The old method works just fine—so why change it?“

… is probably the biggest hurdle we face when introducing new technologies. It’s not the technology itself that holds companies back — it’s our natural resistance to change. Many hesitate to shift away from familiar ways of working, often because they see new systems as disruptive, unnecessary, or at that moment even performing worse. But adopting technology for the sake of novelty isn’t the goal. The real question isn’t whether something is new, but whether it’s better.

The only way to break through this resistance is through results. Time and time again, we must prove that new technologies aren’t about replacing people or cutting corners — they’re about making work more effective, reducing inefficiencies, and allowing teams to focus on what truly matters: serving the public.

Why AI?

If we look beyond the buzzword, AI is really just about using the immense raw computing power of modern computer chips to mimic and automate human activities — whether it’s writing stories, painting pictures, answering customer questions, or making sense of trends and data. The automation makes sense if past events accurately predict future events.

The reason why we can do today what we couldn’t do twenty years ago is the availability of immense, raw computing power. With the help of modern AI chips we can define and describe human activities using billions or trillions of parameters – something impossible by humans. So, no matter how complex we find an activity, if past events accurately describe future behavior, we can use chips to turn the process into a (very big) algorithm, and then leverage the algorithms to automate the process.

Now, imagine applying that same idea to running a business. In theory, any process — from hiring employees to managing inventory to responding to customer inquiries — can be learned and automated by AI. We’re already seeing this in action. Chatbots handle customer service, AI-powered tools draft emails and reports, and smart software helps businesses detect fraud, optimize supply chains, and predict sales trends. But right now, most of these AI tools work in isolation—each one doing its own specific task.

That’s about to change. Over the next couple of years, Agentic AI will evolve from handling individual tasks to managing entire workflows. Instead of just answering customer service chats, it’ll track orders, schedule follow-ups, highlight anomalies, and recommend actions — all without direct human intervention.

Think of it like the early days of factory automation—first, machines replaced individual manual tasks, and eventually, entire assembly lines became automated. This transformation isn’t just about efficiency — it’s about re-imagining how businesses operate. Just like factories in the 20th century scaled up production like never before, AI-driven businesses will be able to scale decision-making, creativity, and problem-solving at levels we’ve never seen. And that’s where things start to get really exciting.

[1] Ford, H. (1930). Moving forward. Doubleday, Doran & Company.