DEV Community: Vincent Tan

Starting Over With the Same Machine

Vincent Tan — Sat, 11 Apr 2026 16:00:54 +0000

The server had been sitting in the corner of the home office for about two months before I did anything with it.

The wind-down was still grinding — creditors, lawyers, suppliers, banks, each moving at their own pace with no regard for mine. But in between calls and emails, I kept finding myself standing in front of that machine. Not staring at it exactly. More like two old colleagues who'd been through something together, neither quite sure what the next chapter looks like.

Same chassis. Same RGB lighting cycling through its colours, completely indifferent to the fact that the company it had powered was gone. The hardware didn't know the difference.

I did. And I had to decide what I actually wanted it to become.

The temptation to just copy-paste

The easy move would have been to carry everything over. The containers were all documented. Spin it back up at home, point the Cloudflare tunnels at the new IP, done in a day.

I nearly did it. Then I looked at what I'd actually be carrying over. Nine months of accumulated services. Containers that had outlived their purpose the moment the office closed. Port numbers that had made sense in November and were confusing by February. A directory structure that had grown organically — which is a polite way of saying it grew without any plan at all.

The staff were gone. The company was gone. Why was I dragging all of their infrastructure into my family home?

So I stopped. And I did something I probably should have done eleven months earlier: I sat down and actually designed the thing before building it.

Designing it properly

One of Claude's more underrated uses isn't writing code or debugging configs. It's thinking. You describe a problem, you push back on each other, and somewhere in the process you arrive at decisions you wouldn't have reached alone.

We looked at homelab forums, self-hosting communities, turnkey platforms promising to do exactly what I wanted. OpenClaw was getting a lot of attention — looked appealing on the surface. The more I read, the less comfortable I felt. Security concerns, oversold features, negative reviews from people who'd actually deployed it. I'd already been through that cycle with AI frontends. I wasn't interested in repeating it.

The conclusion: there was no ready-made solution that did what I wanted. Not because I was doing something exotic, but because what I was building combined things that don't usually go together. Family document management plus password sharing plus AI tools plus a learning environment for a teenager. We drew the map ourselves.

The decisions that mattered

The directory structure got one firm rule: nothing in the root without a reason. Every service, every user, every component gets its place — and if there's nothing in a folder yet, the folder doesn't exist yet.

Port allocation got a proper scheme. The whole home platform lives in the 88xx range, carved into logical blocks — support infrastructure in 880x, app frontends in 881x, MCP servers in 882x, databases in 883x. When you're debugging at 11pm trying to remember which port runs which service, having a scheme you designed yourself is worth more than you'd think.

And then there was Vikunja. I'd been thinking about what it actually meant to build an AI system that could genuinely act on your behalf. For that to work, the AI needs somewhere to put the work. Not just read tasks — own them. Create them, assign them, update them, close them out. Vikunja is a self-hosted task manager — open source, full API, native mobile app. The moment I understood it wasn't just a to-do app but the missing action layer in the agentic model, it stopped feeling like an optional addition.

The spring clean

Before any new build, there was a reckoning with what was already running.

Halfway through the cleanup, I opened the LibreChat config to remove it and ended up reading through what we'd actually built inside it. The Sales Portal. The Workspace Portal. The AutoCount integration. Revenue tracking, inventory analysis, commission calculations, picking list automation — an entire AI-powered operations layer built on top of our accounting software, running silently in a container that nobody was actively using anymore.

I've spent thirty years building things. And in the two months of winding down, I watched all of it get demolished in the way businesses end: not dramatically, just gradually and then completely.

I learned CPR after my mother died so I'd never feel helpless in that moment again. I keep this system for the same reason. The skill exists even when there's no emergency. Someday, if it can help another SME owner — give them the clarity that AI can bring, before it's too late — maybe what I built here means something.

Many containers went. LibreChat stayed. The spring clean was done.

What the home platform became

Eleven containers, each serving a purpose. Document management, password manager, photo library, AI tools accessible from any device in the world. Port ranges documented. Startup chain tested through a real power outage. GitHub repo set up, configs version-controlled, documentation written for a future maintainer who isn't me.

That last part is the important one. Everything I'd built before, I'd built for staff. This was different. I was building for my wife and my son. Building for people who aren't you is a completely different problem. That's what Post 5 is about.

Next: What happens when the people you built something for are too busy to use it.

Why AI Can't Save You From Everything

Vincent Tan — Sat, 28 Mar 2026 16:00:03 +0000

Let me be upfront about something before we get into this.

This is the post I wasn't sure I wanted to write. Not because the story is shameful — it isn't. But because it's the kind of honest that most business owners don't put in public. We're conditioned to share the wins, the pivots, the "lessons learned" wrapped neatly in retrospect. The messy middle is usually kept private.

I'm going to try to do something different here.

2025 was a year I wouldn't wish on anyone

The macro picture first, because it's important context and it wasn't unique to us.

During Covid, the Singapore government did what governments do in a crisis — they made credit available. Schemes, grants, low-interest loans with generous tenures. What nobody fully reckoned with was what would happen five years later when those loans matured and businesses went back to the market for refinancing. Higher interest rates, shorter tenures, smaller amounts. The credit that had been oxygen during the pandemic became expensive and scarce exactly when businesses needed it most.

Then the wars. Ukraine first, then the Middle East. Supply chains that had just recovered from Covid got disrupted again. Goods that used to take six weeks to arrive were suddenly taking twelve. You're paying for inventory sitting on a ship somewhere, and you can't sell what you don't have.

This was the environment most SMEs in Singapore were operating in during 2025. We weren't unique. We were just caught in it like everyone else.

What was happening inside the business

On top of the macro pressure, 2025 decided to get personal.

One of our workers had a heart attack at the workplace and passed away. Someone died. A person who came to work and didn't go home. I was there. I administered CPR, doing everything I could to bring him back — a skill I had learned years earlier, after losing my mother suddenly, because I never wanted to feel helpless in that moment again.

May he rest in peace

I failed.

Beyond the grief, there was an MOM investigation that followed, which is still ongoing.

Meanwhile, my other business — F&B retail — lost two of its best outlets simultaneously. One to competition, one to a temporary closure for renovation. Both at the same time, creating a liquidity drain we hadn't planned for. When the renovated outlet reopened, business was slow to recover — and the landlord raised the rental by twenty percent.

I fractured my right hand in April during a basketball game. Four months largely out of action, trying to run multiple companies one-handed. Literally.

The moment I learnt where the metacarpal is

And in the background, quieter but heavier — family members dealing with cancer diagnoses across two consecutive years. My father, in his mid-seventies and showing early signs of dementia, was giving me all kinds of trouble at work and at home. Then he got into a car accident — the vehicle got scraped but thankfully nobody was hurt — and that became one more thing to sort out on top of everything else.

I'm not listing all of this for sympathy. I'm listing it because this is what "running a business" actually looked like in 2025, behind the LinkedIn posts and the quarterly updates.

What AI actually gave me

Here's the part that matters for this blog.

The AI system we'd built — the dashboards, the live reports, the AutoCount integration — was working. It was doing exactly what I'd hoped it would do. And what it showed me, clearly and early, was a trend I didn't want to see.

Revenue declining. Collections slowing. Costs rising. The gap between them widening in a direction that only goes one way.

Without the system, I might have seen this three to six months later. By which point the options narrow considerably. Instead I saw it while there were still options on the table.

The AI wasn't failing. If anything, it was delivering more clarity than I'd ever had. The plans were good. The direction was right. We just didn't have time.

Time is the one thing no system can manufacture. So I made the call.

Partly because the numbers said to. Partly because I was exhausted in a way that goes deeper than tired. And partly because stopping now — painful as it was — meant I could still do right by the people who depended on me. Pay the staff. Settle with the banks. The people who showed up every day would be taken care of first.

Sometimes the most important thing a good system can do is tell you when to stop.

The server in the corner

While all of this was unfolding, there was a server sitting in the corner of my home office. The same machine that had run the company's operations. Now repurposed. Waiting.

Building the home AI system didn't happen after the dust settled. It happened in the middle of everything. Late nights, early mornings, in between calls with lawyers and emails to creditors. Something to build when everything else felt like it was dismantling. Something constructive in the middle of the chaos.

That story is next.

Next: How a company server became a family AI system — and why building something new was the best thing I could have done.

The Business Build: What We Actually Made

Vincent Tan — Sat, 14 Mar 2026 16:00:50 +0000

Let me set the scene properly.

The server arrived and I was not prepared for what came out of the box. I'd specced it out with AI's help — CPU, RAM, storage, all very sensible and practical. What I hadn't accounted for was that modern PC builds apparently come with RGB lighting. Lots of it. Shifting, cycling, colorful LED lighting behind a transparent side panel that lets you see every single component glowing like a nightclub.

In the office.

On my desk. Because I didn't have a separate workstation, so the server was just... sitting there. In front of me. Looking festive. With staff throwing me that weird look.

Thank god the see-through panel is on the right!

That was day one.

What we were actually dealing with

I was managing a few companies, handling thirty to forty staff. Like most small businesses in Singapore, we were held together with a combination of determination, muscle memory, and software that had accumulated over the years rather than been deliberately chosen.

The file server was a good example. It had been running since 2003 — old enough to vote — and had grown organically ever since. Inside it we found documents dating back to 2013, buried under layers of folders that made sense to whoever created them and nobody else. Multiple copies of the same file with no clear indication of which was current, which was stale, or why three versions existed in the first place. Finding something was a coin flip. Reports took days to compile because you had to pull data from three different systems, reconcile them manually, and hope nothing had changed by the time you finished.

And the physical documents. Arch folders in cabinets, cabinets spanning half the office floor. Every time someone needed something from that archive, it was an expedition.

The trap I walked into

Here's something nobody warned me about when I started planning all this with AI's help.

On paper, everything sounds straightforward. You describe your problem to an AI, it maps out a solution, the solution sounds logical and achievable. What the AI doesn't tell you — what I now call the digital fine print — is everything that sits between the clean description and the actual working reality. The legacy software that turned out not to be compatible with Windows Server 2025. The recommended tools that looked great in documentation but felt clunky the moment real users touched them. The race conditions, the configuration gotchas, the tools that had been superseded by something better but the AI's training data hadn't caught up yet.

We went through three AI chat platforms before we found one that stuck. AnythingLLM, then Open WebUI, then LibreChat. Each switch felt like a setback at the time. Looking back, it was just learning.

And then there were the ideas that were genuinely too ambitious. At one point I was seriously researching fine-tuning a language model on my own conversations — the idea being that I'd leave behind not just a document archive but a digital version of me. A legacy AI. I still think that's a fascinating idea. It was also wildly beyond what I should have been attempting at that stage.

The AI's ugly side

This is the part the productivity influencers don't show you.

AI hallucinates A LOT. It confidently tells you to do something that turns out to be wrong, or outdated, or technically correct but catastrophically applied to your specific situation. I lost months of container data once because of a configuration change that seemed reasonable and wasn't. I spent entire evenings untangling problems that an AI had talked me into creating in the first place.

When the RTX 4000 Pro finally arrived, my first reaction was mild disappointment. I'd been watching YouTube videos of people unboxing 4090s — massive, triple-fan cards. The Blackwell professional card is nothing like that. Slim, understated, almost boring-looking.

Not the GPU you imagined it to be

It took me a moment to reconcile "this quiet little thing" with "24GB of serious AI compute." Turns out the ones doing the real work don't need to show off.

What we built that actually worked

Thin clients. When a machine broke down, instead of spending thousands on a new PC, we replaced it with a thin client at a fraction of the cost. One thing I discovered: some modern smartphones have a desktop interface mode — connect them to a monitor and keyboard and they work like a thin client.

The document archive. Painful to build. But once stable — genuinely one of the most satisfying things I've built. Every document searchable, tagged, retrievable in seconds. Information on demand, instead of information on expedition.

The AutoCount integration. Our accounting software now talks to everything else. Sales tracking, procurement analysis, commission calculations that used to require hours of manual spreadsheet work now run in seconds.

Purchase tool built with Claude

The picking list. What used to require three people to manually compile and verify now runs with one. The error rate dropped to near zero.

The results

The SaaS contracts we dropped, combined with ending the IT support retainer, freed up thousands a month. For a company our size, this wasn't incremental improvement.

Sales Dashboard completed

It was a genuine step change.

So why did it wind down anyway?

If you built all this, if the savings were real, if the operations improved — why are you winding down?

Because AI can fix a lot of things. It turned out there were things it couldn't fix.

That's Post 3.

Next: What I learned about the limits of technology when the real problems aren't technical.

When most people hear "home AI system", they think smart lights and Alexa. Mine started with grief.

Vincent Tan — Thu, 26 Feb 2026 19:23:15 +0000

I'll be honest — I was late to the AI party, and when I finally showed up, I wasn't impressed.

ChatGPT was everywhere in early 2024. Everyone was talking about it. So I tried it. And yes, it was clever. But after a while, it just felt like a very confident search engine that occasionally made things up. I got bored. I moved on.

Then Gemini caught my attention. And this time, something clicked.

I was dealing with a problem at work — a SaaS vendor charging obscene amounts for a service that frankly wasn't worth it. I'd been putting off doing anything about it because the research alone felt overwhelming. Where do you even start? What are the alternatives? What would migration look like?

So one evening I just... started talking to Gemini about it.

A few hours later, I had a clearer picture of the problem than I'd had after months of procrastinating. Not because the AI solved it — but because it cut through the part that always stopped me: the research. The dead ends. The time spent just trying to understand the landscape before you can even think about solutions.

That was the moment I understood what AI actually was. Not a chatbot. An enabler.

And once I saw it that way, I couldn't stop.

Three problems. One probably-too-ambitious plan.

Here's where I have to admit something: I tend to think big. Sometimes too big.

At the time, I had three problems sitting on my desk that had been there for years:

Problem 1: The company server was dying. Seven-plus years old, held together with prayers and deferred maintenance. It needed replacing. But if I was going to replace it, I wanted to do it properly — not just buy another box and repeat the same mistakes.

Problem 2: I wanted a proper family system. Passwords, documents, photos, videos — all scattered, all fragile. But this wasn't just a convenience thing for me. It was personal.

My mother passed away suddenly from a heart attack. No warning. No time to say goodbye, and no time to leave instructions. On top of the grief — and grief is already brutal — we had to scramble through her affairs trying to piece together her estate. What accounts did she have? Where were the documents? What did she want?

I remember thinking: there has to be a better way. Not to avoid death — nothing avoids that — but to make sure that when it happens, the people left behind aren't drowning in chaos on top of everything else. A proper family archive. A legacy system. Something that actually preserves what matters.

Problem 3: I was done with SaaS that didn't fit and IT support that couldn't keep up. SaaS sounds great in theory — someone else handles the infrastructure, you just pay and use. The reality is that SaaS is designed to serve everyone, which means it ends up fitting no one perfectly. You pay for a hundred features, use ten, and spend half your time working around the ones that almost-but-don't-quite match how your business actually works.

The IT support wasn't bad people — it was a structural problem. One technician covering thirty to fifty companies. You do the math. By the time they got to your problem, you'd either solved it yourself or given up. And every hour they spent on you was billed accordingly.

So instead of solving these problems one at a time, like a sensible person, I decided to solve all three simultaneously. With a server I'd never built before. Using AI tools I'd just discovered. With IT knowledge that peaked sometime around DOS and a 386 processor.

Yes. I know.

The Frankenstein server

I started researching parts. AI helped me understand what I needed — CPU, RAM, storage, networking. But the GPU was the headache.

This was mid-2024. The AI hype had consumed every decent GPU on the market. I wanted a proper one for running local AI models, but the wait was real. We ended up running on a loaner GPU from the vendor from April all the way to November. Seven months of "we're getting there."

When the RTX Pro 4000 finally arrived in November, it felt like Christmas.

In the meantime, I was learning. Slowly. Painfully. Wonderfully.

The early days looked like this: open terminal, run a command, get an error, copy the entire output, paste it into the AI chat, read the explanation, try again. Repeat. For hours. For weeks.

It was not fast. It was not elegant. But every small win felt enormous, because I was doing things I genuinely didn't think I could do.

The moment everything accelerated

Somewhere around May, I discovered Claude.

I don't remember exactly what made me switch. But I remember the moment I first used the Artifacts feature — the ability to generate and iterate on actual working files, not just text — and something shifted. It felt like that scene in a movie where the actor stands still while the world transforms around them. Technology moving faster than you can track. Every week, something new.

MCP — Model Context Protocol — was the real inflection point. Suddenly AI wasn't just answering questions. It was connecting to systems, taking actions, building things. An API integration that used to take me days, sometimes weeks of research and trial and error, could now be scaffolded in hours.

I had to stop and ask myself: what am I actually building here?

The hard lesson

The honest answer was: too much.

I'd set out to solve the family system AND the company infrastructure AND replace the SaaS stack AND build an AI platform, all at once. For someone whose last serious IT experience was watching Windows 95 install from floppy disks, this was ambitious to the point of delusion.

So I made a call. Family system goes on the backlog. Focus on the company first. Get the business results first, and let the personal project wait.

It was the right decision.

Over the following months, we replaced our aging PCs with thin clients. We built a full digital document archive with proper ingestion and retrieval — no more hunting through filing cabinets. We connected directly to our accounting software's database for real-time analysis. We automated our delivery picking lists — what used to require three people checking and cross-referencing orders now runs with one, and the error rate dropped to near zero.

We dropped several SaaS subscriptions. We parted ways with our IT support company. The savings were significant — thousands a month.

And I did most of it myself. With AI as the co-pilot.

So what is this blog?

I'm Vincent. I'm 50. I ran a trading company for years and I'm in the process of winding it down. I'm not a developer. I'm not a cloud architect. My formal IT education stopped somewhere around the age when owning a CD-ROM drive made you cool.

But over the past eleven months, I've actually built two production-grade AI infrastructures. The first for the company — the one that cut costs, automated the picking lists, replaced the SaaS stack. Then the company wound down, the server came home, and within about a week the whole thing was repurposed into what it is today: a family AI system running our documents, photos, passwords, and memories.

And I want to write about it. Not to show off. But because I think a lot of people like me — business owners, working professionals, people who are curious but intimidated — are still using AI as a smarter Google. Ask a question, get an answer, close the tab, move on. And walking away thinking: okay, neat trick, but not really for me.

I thought that too. For longer than I'd like to admit.

This blog is about what changed my mind. And how you can get there faster than I did.

Next: What exactly did I build — and what does a "family AI system" actually mean when it's not about smart lights?