Home safety is something I was raised to take seriously. I always found myself checking that doors are locked before bed, being aware of unfamiliar sounds, and generally having a plan in case something felt off. That mindset has stayed with me, but the tools I use today look very different.

Recently, I built a small home lab that acts as the central hub for what I think of as home tech defense. It integrates environmental sensors, lighting, file storage, network security tools, and security cameras into a single system that I can monitor and control locally.

The sensor setup is intentionally broad. I currently monitor temperature in every bedroom and the living room, CO₂ levels, cooking gas, and particulate matter (PM1.0, PM2.5, and PM10). I also use simple door sensors to detect whether entry points are open or closed.

These sensors serve two purposes. First, I strongly believe in monitoring indoor air quality—especially in tightly sealed apartments. Second, and more honestly, they compensate for human error. I’m forgetful. Sometimes I worry that I left the gas on or didn’t fully close the door before leaving home. Being able to check sensor data remotely gives me immediate confirmation and significantly reduces anxiety. In almost every case, the alert turns out to be a false alarm—but the ability to verify matters.

What’s notable is how inexpensive this setup is. All of the sensors combined cost me under $100 in parts. I used ESP32-based boards with Zigbee capability, along with an SCD40 for gas sensing, a BMP280 for temperature, a reed switch for door state detection, and a PMS5003 for particulate matter. Everything required manual soldering and integration, but the result is a flexible, privacy-preserving system that runs locally.

Another core component of this home defense setup is camera surveillance software. I use Frigate to manage security camera feeds and perform on-device AI inference. This enables object detection, motion classification, and even license plate recognition without sending video streams to the cloud. For systems without a dedicated GPU, adding a Google Coral TPU provides efficient, low-power AI inference at the edge.

Looking ahead, one idea I find particularly interesting is the use of small, autonomous—or manually controlled—indoor drones as part of a home defense system. Imagine being home alone late at night and wanting to verify that no one else is in the apartment. Instead of physically checking each room, you could deploy a small drone from the safety of your bedroom and inspect the space remotely. In a more advanced setup, the drone could even be triggered automatically by sensor events and predefined rules.

This isn’t about turning a home into a fortress. It’s about using affordable hardware, local AI, and automation to extend awareness, reduce anxiety, and give people better control over their own environments.

I’m curious which part of this would be most useful to people: environmental sensors, local AI camera analysis, or automated inspection tools. Send me an email!

The United States has a spending problem—not just at the government level, but among individual households as well. Americans collectively hold over $1 trillion in credit card debt, and in a recent survey of 1,000 Americans, roughly one-third reported being maxed out on their credit cards. Overspending and undersaving aren’t just math problems; they’re deeply psychological ones. (I highly recommend The Psychology of Money for anyone interested in this topic.)

Tracking personal finances is harder than it sounds, especially when you’re busy or raising a family. For many people, “budgeting” ends up being a rough estimate—spending money and hoping the total stays below the next paycheck. Even for someone who pays attention, it’s difficult to keep track of every credit card purchase, watch for duplicate charges, or notice subtle fraudulent activity.

There are also broader financial decisions that most people know they should optimize but rarely do: shopping around for better car insurance every six months, understanding whether their insurance coverage is actually appropriate, or evaluating whether the terms in a property purchase contract are fair. These are high-impact decisions, and they’re exactly the kinds of problems where AI can be genuinely useful.

To explore this, I started building a personal finance application for my own use. The first capability I added was simple: detecting duplicate charges. I connected my bank and credit card accounts, downloaded my transaction history, and used a local Qwen-3 model with a custom prompt to analyze the data. The model flagged potential duplicates quickly, saving me the manual effort of scanning through hundreds of transactions.

Next, I wanted to see whether AI could surface anything of value in my spending patterns—not just errors, but insights. I wrote a second prompt designed to look for anomalies, unusually high recurring expenses, or optimization opportunities. In my case, my finances were generally in good shape, but the model did identify one useful suggestion: my car insurance premium was slightly high compared to similar spending profiles, and I should consider shopping around.

The most impactful use case, however, came when I purchased my apartment. At the time, I was busy with work and other life commitments, yet suddenly faced a stack of legal agreements, disclosures, and contracts that I was expected to review and sign within days. Before large language models, I likely would have skimmed these documents—or not read them at all.

Instead, I fed the contracts into an LLM and asked it to explain the terms, highlight risks, and point out areas that might be negotiable. To my surprise, the analysis was genuinely helpful. It identified issues I would have missed and gave me the confidence to push back. As a result, I saved a few thousand dollars, had several repairs addressed, and even received a cash credit for a non-functional refrigerator that the seller agreed to cover.

This experience made something clear to me: many personal finance mistakes aren’t caused by irresponsibility, but by cognitive overload, time pressure, and complexity. AI doesn’t replace judgment—but it can dramatically reduce the cost of paying attention.

---

AI has evolved rapidly over the past few years, with new models, architectures, frameworks, and open-source tools making it possible to tackle workloads that were previously out of reach. What’s changed most is not just capability, but accessibility—AI is now something individuals can experiment with, adapt, and build on directly.

This blog is where I’m documenting that journey, with a focus on creating practical AI applications that solve real problems. I’m especially interested in how AI can be used in everyday areas like personal finance, understanding complex legal documents, and automating decision-making.

In future posts, I’ll explore topics such as AI prototypes, agentic AI workflows, and retrieval-based systems, as well as how to run AI models locally on a laptop. Running models locally avoids dependence on API keys and keeps sensitive data private, which I see as an increasingly important part of modern AI use.

AI has come a long way—and there’s still a lot to test, question, and build. Let’s dive in.