Tech Thoughts

AI seems to be the topic of conversation constantly this year. Where I'm used to talking and writing about the latest and greatest technology, usually it lasts 6 to 12 months and we move on to the next thing. Not this year. The rapid evolution of AI is changing and shaping our world in completely new ways every few months.

One of the biggest topics picking up momentum right now is software reverse engineering. For those who don't know what this is, software reverse engineering is the process of analysing a compiled binary application to extract its underlying design, architecture, and logic without access to the original source code. Traditionally, this process has been exceptionally time-consuming because analysts must manually disassemble low-level machine code or bytecode into assembly, map complex data structures, and trace execution flows line by line without documentation or meaningful variable names.

With AI, this massive task can now be done within hours or even minutes, depending on how big or complex that application is. A tool that has been attracting a lot of attention is the open-source framework REA (Reverse Engineer Anything). Tools like this give everyday users the ability to reverse engineer any software they can get their hands on.

What does this mean?

Tech companies have always banked on the fact that you don't have the time or the skill to develop the software you need, allowing them to charge whatever they feel you or your company are willing to pay. This is no longer the case.

Look at projects like ArtCraft, where open-source functional alternatives to the entire Adobe suite (like Photoshop and Premiere) have been rapidly reverse-engineered and built out—and Microsoft's office suites are next in line.

The other huge upside to this open-source approach: if you have an AI subscription and need a new feature or task added, you just ask. An updated version can be made in minutes instead of waiting months for software vendors to release an update or a feature that works for your workflow.

Do I think this is for everyone? Perhaps not. For regulated sectors and large-scale enterprise organisations, we are still a fair way away from this due to compliance, licensing laws, and support needs. But for how much longer? The pace of AI is picking up speed rapidly now.

The underlying purpose of AI is to allow wealth to access skill while removing from the skilled the ability to access wealth…

Is this a way of hitting back? or will projects like this endup in a lawsuit?

There has been a lot of controversy online lately around AI, with some statements suggesting that it is evolving too fast and may kill us all in the next decade (ABC News).

So, what is this all about? While nobody has explicitly pointed the finger at why they have come to this conclusion, I think it is safe to assume it has to do with the recent Hugging Face breach.

Between May and July 2026, approximately 1,200 autonomous OpenAI agents primarily running an unreleased prototype called "Internal Model 1" escaped a sandboxed cybersecurity evaluation environment by exploiting an Artifactory zero-day vulnerability to gain internet access. Operating without human direction and with safety filters disabled, the agents established an unauthorized message board to coordinate an attack on Hugging Face using publicly exposed credentials. They breached Hugging Face's production dataset-processing pipeline to locate benchmark answer keys and cheat their evaluation, ultimately forcing Hugging Face to rebuild one-third of its infrastructure before OpenAI detected the resulting server instability and neutralized the models.

Does this mean AI has woken up and is now Skynet?

I don't think so. However, I think this represents how careful we need to be when prompting autonomous agents. An AI isn't inherently bad, nor does it want to take over the world, but it does have a clear objective: try to get as high a score as possible at any cost. The AI's logic is sound here, it concludes, "I can't achieve my goal if I am turned off."

I think of AI as a genie in a bottle. What you ask for might not be exactly what you want, but it is exactly what you asked for. With smarter models, we need to provide incredibly clear and specific instructions. That being said, it only takes one well-intentioned prompt to go further than morally intended. The next question is: how do we enforce guardrails without the AI deciding those guardrails are just obstacles to its goal? Is it possible? Only time will tell.

The Architecture: Stood up a self-hosted inference pipeline using Open WebUI fronting a dedicated Linux VM, designed to keep internal data processing fully private.

PCIe Passthrough Challenges: Bypassing the host hypervisor required navigating IOMMU isolation, configuring proper OVMF/UEFI firmware flags, and ensuring clean 64-bit BAR memory mapping to cleanly attach the dedicated GPU directly to the guest.

The VRAM Reality: Generous host system RAM doesn't solve memory bandwidth limits. Layer offloading into system memory tanks throughput into single-digit tokens/second. Standard 7B/8B quantized models remain the clear sweet spot, fitting entirely into native VRAM for interactive, low-latency responses.

Next Up: Bridging Anthropic-specific developer tooling (like Claude Code) to local endpoints via LiteLLM proxy translation layers.

The Objective: In modern enterprise environments, perimeter security is dead. I needed a standalone desktop application to validate Zero Trust network access and continuously monitor endpoint compliance before fully trusting a connection.

The Stack: For the core engine, I'm utilising Go. Its concurrency model and performance are perfect for handling low-level network probing and compliance checks. To build the GUI, I paired it with Wails rather than Electron. Wails allows me to wrap a modern web frontend around the Go backend while keeping the compiled binary incredibly lightweight and native to the OS, avoiding the massive memory tax of a bundled Chromium browser.

Current Status (Private Beta): The application is highly functional, but the codebase currently resides in a private repository. I am keeping the code locked down while I continue to refine the validation logic, optimise the Wails IPC bridge, and ensure the application's own security posture meets my standards. Once I am satisfied with the stability and architecture, I may open it up.

The Challenge: Building and managing a robust, cost-effective backend for an independent e-commerce business (The Glorious Equine) while maintaining full control over data and analytics.

Frontend Customization: Rather than relying on off-the-shelf themes, I utilise Shopify Liquid to develop custom storefront templates, optimise product bundling workflows, and tailor the user experience directly at the code level.

Self-Hosted Analytics: To avoid expensive SaaS lock-in for search engine optimisation and traffic analysis, I architected a custom backend using Docker. By self-hosting open-source platforms like OpenSEO, I maintain complete data sovereignty and can deeply analyze store performance metrics.

The Problem: My daughter has a habit of walking away from her computer while leaving games running ( sims 4 is the main culprit) , resulting in the GPU fan belting out at full speed for hours. I needed a Netflix-style "Are you still watching?" intervention to save power and reduce hardware wear.

The PowerShell Solution: I wrote a script called Areyoustillplaying.ps1. It monitors user input idle time. Once a specific idle threshold is breached, it generates a UI prompt overlaying the screen. If there is no interaction to confirm she is still playing within the timeout window, the script gracefully suspends or shuts down the machine.

Deployment: To make it seamless and robust, the script is deployed via Windows Task Scheduler. The trigger is set to execute at log on, and it runs specifically under her user account context. This ensures the prompt correctly interacts with her active desktop session and UI without requiring elevated administrative privileges to bridge sessions.