The 7th lunar month has arrived. In Vietnam, China, and other cultures influenced by Chinese traditions, people believe this is the time when the gates of hell open, allowing hungry ghosts to wander the human world in search of food.

This month often features a unique activity called "Giựt cô hồn" (snatching the ghost's offerings). It happens when working-class people or children rush to grab offerings—like candies, small change, and fruits—from the ritual tray dedicated to wandering souls. This takes place right after a household finishes their ghost-feeding ritual. It symbolizes sharing blessings and warding off bad luck.

To me, this is a very familiar image because I used to join in on these activities during my childhood. I've noticed that kids nowadays don't know much about "Giựt cô hồn" anymore. Perhaps it's because life is more comfortable now, and the newer generation simply doesn't care as much about these kinds of traditional rituals.

Just a little note to myself for today.

A person can learn many technical skills or management theories in a short amount of time. But to achieve a truly relaxed, forgiving, and free state of mind towards past pain or people from the past—that is the result of a long process of inner practice and profound awakening.

Da Lat is probably the place I've visited the most in 2026. Even though it's 400km from my house, I've already been there twice in just the first four months of the year. And there's a pretty high chance I'll make another trip for the September 2nd holiday, since several groups of friends have invited me.

Honestly, Da Lat isn't as beautiful to me now as it was when I first visited at nine years old. The reason is simple: it's become heavily concretized, almost to the point of feeling suffocating. It's also overcrowded with tourists, making it hard to find any real quietness anymore.

However, the cool climate—sometimes dropping to 9 or 10 degrees—still holds a certain appeal for someone like me who loves cold weather.

I've never tried staying in Da Lat for a long time. Usually, by the third day, I'm ready to go home. It's not exactly that I get bored. It's just that if all I'm doing is eating and relaxing, by day three I start missing the feeling of being busy.

Many people from Saigon have moved to Da Lat and shared interesting experiences about living there. I don't think I'll do the same. But finding a job in Da Lat when I no longer have to worry about finances and kids might not be a bad idea.

Only time will tell.

I still have days where my mind just runs in loops, worrying about everything.

The recent news about university entrance exams is rough. In some majors, kids need near-perfect scores just to have a shot at getting in. But then you look at the other side—high unemployment rates for fresh graduates and the constant, looming threat of layoffs. It just feels so chaotic for the younger generation right now.

And honestly, it's chaotic for me, too. I'm caught in this weird middle ground, trying to survive as an employee while also being a parent. I find myself worrying about where my own career is heading, while simultaneously stressing out about what kind of future my kid is walking into.

For now, I'm trying to figure out how to protect my own livelihood and build a few solid backup plans. But maybe the bigger challenge is figuring out how to guide my son. I need to make sure he picks up the kind of skills that will actually help him navigate a world that feels completely unpredictable.

When I was younger, before I figured out I wanted to be a software engineer, I always thought I would become a designer.

But life took many different turns. I became a software engineer, and gradually, I completely abandoned my drawing skills.

It has been over 20 years since I last held a pen to draw. My hands are shaky now, and controlling the strokes feels quite difficult. Still, I recently tried sketching again, and to my surprise, the result wasn't too bad.

I recently added a comment feature to my blog. The goal isn't to build a public forum, but to give myself a way to revisit old posts and add new notes or reflections over time.

I thought about opening the comments for everyone, but that feels contradictory to why I built this space. I want to avoid the noise, notifications, and endless feedback loops of the internet and social media.

This blog remains my quiet corner. Still, looking back at old entries and appending fresh thoughts as a simple comment is an interesting way to track how my perspective evolves.

My initial goal was just to lose weight and reduce visceral fat to help treat my prediabetes. After successfully reversing my prediabetes, I got used to the training intensity and have now taken it to a higher level.

I have started buying gear like knee pads, and I am considering getting a pair of specialized running shoes. My current distance is up to 8km in a single run. I can run continuously for over an hour—something I thought was impossible just six months ago.

My current fitness level is more than enough to conquer 10km or even 15km milestones. As long as I keep my pace between 8.5 and 9.5, I can easily extend my running distance.

My weight has stabilized around 63 - 64kg. However, I need to rethink my nutrition plan to better suit my current needs. I am now a healthy person, not the sick patient I was three months ago.

Swarm (formerly Foursquare) is an app I have been using since 2010. It is a location-based social network. Back then, I worked for a startup in this space, building a similar application. That is why I kept using it long after I stopped working there.

Working on that app helped me discover interesting places and great food spots in Saigon, as well as many other places I visited.

Foursquare as we knew it is gone, succeeded by Swarm. I still use it today simply for the sake of my youthful memories. It holds onto the photos and the places I have been.

I have a history of building products by myself, only to fail when it comes to operations. I've also started many things, only to quickly get bored and give up.

But with killedbyai.xyz, things are different. I have made a serious investment in this project.

What started as a personal "just for fun" project has evolved. I began viewing a data-driven approach as the key to improvement. By actively tracking metrics from Google Analytics and listening to user feedback, I found a clear path to make the product better.

It has been 5 months since the first version. Today, I deployed a brand new UI for the website. This update moves the project in a more professional direction. It is the beginning of going deeper—turning it into something much more than just a website that provides information.

I have been rereading "The Peter Principle" recently, and each time, it gives me new insights into my work managing teams. I am not surprised to see the ideas in the book play out in my daily work.

What I wonder about is this: if "The Peter Principle" explains the problem so clearly, why do companies keep making the same classic management mistakes? Why do we still promote the wrong people or force great Individual Contributors (ICs) onto the management track?

In a large organization, it takes a long time to realize that a manager is struggling and has hit their ceiling of competence.

Companies don't promote the wrong people because they are unaware of the Peter Principle. They do it because fixing the real problem is hard. Changing a company's reward structure and shifting how we evaluate people requires a lot of courage and comes at a high cost.

Most organizations choose the easy path for the short term. The price they pay is a stagnant system in the long run.

This year, for my 44th birthday, I decided to do a complete refactor—not of a codebase, but of my own health and body.

For a long time, it was easy to let physical activity slide. But as the years go by, you realize a hard truth: health is the most important thing we have. This isn't just about physical fitness; it is deeply tied to mental health. I have officially brought daily movement into my life. It is now a core part of my routine, as essential as breathing or writing code.

Beyond the physical changes, this year has been a deep dive into understanding psychological issues and how they impact daily life. We often spend so much energy trying to understand other people, trying to figure out why they do what they do. But I've learned that this is mostly pointless.

The real work is understanding yourself.

By turning the focus inward, I've started to recognize my own boundaries. I am learning to accept my own limitations instead of fighting them. There is a profound sense of peace that comes when you stop trying to be everything to everyone and just accept who you are.

Here is to 44: moving more, understanding others less, and understanding myself better.

After dealing with serious FOMO from reading endless articles about Obsidian and Claude, I've finally found my own use case for it.

The difference for me is pairing it with the Antigravity CLI. Now, I truly see the power of Obsidian as a daily assistant for researching new technologies and summarizing my day's work.

There is probably still a lot more to discover within Obsidian's ecosystem, but at the very least, I finally understand why people praise this app so highly.

I just finished experimenting with a RAG-based chat system powered by llama.cpp and the Hermes-3-Llama-3.1-8B-GGUF (Q4_K_M) model.

Spinning up the local inference server required a few specific flags to ensure compatibility with the langchaingo orchestrator—specifically enabling embeddings and configuring mean pooling:

llama-server -hf NousResearch/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M --port 8080 -c 8192 --embeddings --pooling mean

It was quite a journey. I started with basic similarity searches against a pgvector database and eventually had to implement a ReAct loop to make the chatbot respond more intelligently. The difference in output quality was significant.

The system performs surprisingly well when running locally. However, the thought of deploying this as a multi-user production system is a completely different story. The operational costs and infrastructure headaches to scale LLM inference are massive.

Regardless of the deployment challenges, building this small RAG system from scratch was incredibly valuable. It gave me a much more grounded, realistic perspective on the day-to-day realities of being an AI engineer.

Feature speculation in a vacuum is engineering tech debt waiting to happen.

You cannot build software based on subjective assumptions or isolated developer intuition. The only accurate validator for what stays, what gets refined, and what gets pruned is real production runtime.

Deploying to real users under actual operational conditions immediately exposes system realities:

• Unnecessary features become obvious bloat.
• Critical UX gaps surface instantly.
• Architectural assumptions meet hard data.

Across multiple enterprises, I’ve repeatedly seen the same systemic failure: a total disconnect between Product strategy and Technical execution. Few organizations truly acknowledge or actively close the Product-Tech gap. Features are shipped to meet arbitrary roadmaps rather than validated runtime feedback.

Moving forward, closing this gap is the core directive for myself and my engineering teams.

Stop guessing in staging. Ship to production, observe real user telemetry, prune the noise, and let hard operational data dictate the refactoring pipeline.

The paradigm of AI engineering has rapidly matured over a short timeline, moving from basic text manipulation to complex systems engineering:

• Prompt Engineering (2022): Crafting isolated, single-exchange instructions for a model.
• Context Engineering (2025): Managing state, memory, and structural inputs beyond the initial prompt window.
• Harness Engineering (Feb 2026): Establishing strict execution constraints and guardrails to keep agentic behaviors bounded.
• Loop Engineering (June 2026): Designing autonomous plan-execute-verify iteration cycles.
• Graph Engineering (July 2026): Interconnecting multiple agent loops via a shared, synchronized state.

The Engineering Takeaway

We have officially moved past the era of treating LLMs as standalone conversational interfaces. AI engineering is no longer about crafting clever phrasing; it is purely about architectural orchestration.

Single-loop agents inevitably hit execution plateaus. Scaling agentic capabilities demands explicit state machines, controlled boundary enforcement (Harnessing), and structured topology (Graphing).

If you are still optimizing raw prompts in isolation, you are optimizing legacy code. True technical leverage today sits at the orchestration layer—wiring specialized multi-agent loops into deterministic, state-driven execution graphs.

This is the official test run—the very first microblog post generated, formatted, and pushed directly to aslando.life using my new AI-powered Telegram command plane.

Zero code editors opened. Zero manual git commits. Zero friction.

I captured the core thought, dispatched it to my domain-specific writing agent via Telegram, and let the automated pipeline handle tone alignment, frontmatter generation, and deployment.

The control plane is live. The workflow is streamlined. High-throughput publishing starts now.

I’m deploying a small but high-leverage architectural upgrade to aslando.life: an AI-native Telegram bot that serves as my primary control plane.

Currently, the bot handles basic telemetry—delivering real-time engagement metrics, site analytics, and runtime health updates directly to my phone. But monitoring is just Phase 1.

The ultimate goal is to evolve this bot into a domain-specific writing agent. I am training it to deeply understand the "Aslando" voice—sharp, direct, system-oriented, and concise.

Once fully integrated into my publishing pipeline, this agent will allow me to draft, format, and push microblog posts directly via Telegram commands.

Instead of opening an editor, writing frontmatter manually, and running git commands on a laptop, I can capture an observation in the field, run it through my fine-tuned agent, and publish it instantly.

A personal website shouldn't just be a static repository. By wrapping it with a dedicated AI agent, aslando.life transforms from a passive blog into an interactive, zero-friction extension of my own thinking.

As part of the ongoing maintenance and expansion of killedbyai.xyz, I’ve pushed our underlying dataset onto Kaggle: AI Job Displacement & Tech Exposure Insights.

Beyond raw software engineering, I am using Kaggle as a playground to execute end-to-end Data Analytics workflows—the kind that directly inform strategic decision-making.

Instead of just maintaining application code, I am building exploratory notebooks around these datasets to extract actionable patterns: mapping AI task exposure, analyzing industry vulnerability, and visualizing structural shifts in technical labor.

Code handles execution, but clean data drives direction.

Shipping features and maintaining pipelines is only half the battle. By transitioning from pure code implementation to structured data analytics, I’m turning killedbyai.xyz from a passive tracking project into a data-driven tool backed by hard evidence.

After months of strict discipline to bring every biological metric back to green, I had my first re-entry session: draft beer (bia hơi) and good food with colleagues.

The experience revealed a permanent shift in my internal architecture.

In the past, I could easily consume over 4 liters of beer in a single sitting. Now, my system maxes out at less than a single liter. The food was still delicious, but the psychological thrill—the cheap dopamine fix once attached to alcohol and heavy meals—has completely vanished.

My primary drivers have been remapped. Real satisfaction no longer comes from consumption, but from continuous technical discovery, skill refinement, and physical output.

Case in point: the morning after taking in a non-trivial amount of alcohol, I executed a uninterrupted 4km run at a solid 6–7 pace. No lingering sluggishness, no broken routines—just execution.

Restoring physical metrics through diet and exercise was merely the surface-level patch. The real outcome of the last three months is a total refactoring of my operating system: my mindset, my reward mechanisms, and my tolerance for low-leverage habits.

When you upgrade your mental firmware, legacy vices lose their runtime execution power.

I’m currently pushing a deep refactoring cycle for killedbyai.xyz, focusing on fine-tuning specialized skill definitions for my custom agent suite.

To improve overall output fidelity, I’ve moved away from single-model dependency. While Gemini handles high-throughput synthesis exceptionally well, pairing it with Claude for complex reasoning and nuanced generation creates a far more resilient pipeline.

However, scaling a multi-agent system introduces an immediate operational threat: uncontrolled autonomous drift.

After letting agents run with too much autonomy for a period, output predictability degraded. Unmonitored agentic execution without explicit boundaries quickly leads to garbage output and wasted tokens. To fix this runtime anti-pattern, I transitioned orchestration over to Orca.

Orca provides the explicit control plane needed to direct, manage, and audit multi-agent workflows. It enforces deterministic constraints over agent communication while still leveraging their parallel problem-solving capabilities.

For pure multi-agent coordination, Orca is rapidly becoming my primary driver.

AI agents are only as good as the system architecture governing them. Giving models infinite freedom sounds good on paper, but engineering excellence demands strict orchestration, explicit boundaries, and hard metrics.