I've just build a SSD builder to anyone who would like to try it out. This builder is designed according to how I've facilitated a Scenario-System Design session.
Starting with scenarios, my team would always throw in all sorts of ideas of what users might use our features for. If we are implementing a payment gateway, we would immediately indicate Visa and Mastercard as this is our business requirement. Then we would also start brainstorming other gateways that our stakeholders didn't raise up, such as Nets, PayNow, PromptPay, YouTrip, and many others. At this stage, we are not judging the scenario yet. As long as it is a valid scenario, it deserves a spot.
After we have thrown in every scenario we can think of, then we start tagging whether the scenario is probable or an edge case. We might also rank the scenario in specific sequence to indicate high or low likelihood of the scenario coming to pass. This step is optional in most cases. Even my team don't always sequence the likelihood.
At the next step, we would design the system based on the scenarios we listed before. Obviously, the scenarios marked as "immediate" holds the most weight and the design would need to fulfil that scenario. Followed by the "probable" scenarios. Usually we would ensure the probable scenarios are accounted for, although if the team feels that it can be excluded, it's allowed to be excluded. It could be for reasons like a revamp is scheduled within the next 12 months so it's not beneficial to plan for it now. Then the "edge case" scenarios, which is almost always excluded. It's typically added only because in the off chance that it actually happens and it could cause a revamp of the feature, then it's better to consider it now.
And finally, we discuss what we will actually implement in this run. We don't waste time implement something that we ain't going to need. Code is liability. Everything code written and every resource deployed is a liability the team must maintain. We don't want to implement anything not necessary.
Before the age of AI, everything that we didn't need to build helps us deploy faster and every feature we need to build later without a revamp is a business edge. But now that we have AI writing codes at inhumane speeds, it doesn't mean we can build every dang feature our brilliant minds can think of. Every code written and every resource deployed is still a liability that AI must maintain. Liability translates to costs. Either in tokenomics or actual physical server resources.
So with this builder, I hope this aids anybody going through this 3 step process of Scenario-System Design. I'm going to use it myself and refine it further.
I've also added privacy policy, in case you're concerned about the data you put in there. I don't collect any information and I don't intend to. The site is hosted on S3 anyway, so there's no processing facility on the server to store your data. It's all on your browser. The only data I collect, with your consent, is the usage itself. Feel free to opt out of its usage if you're also concerned about it. But if you would be so nice to help me, I would be really grateful to understand your usage, so that I can improve on it.
Now I'm thinking of what feature to work on next. I'm thinking of adding some AI capabilities, such that you can have a conversation with the cards you add, brainstorm together with AI, and manipulate the cards on your behalf. What do you think? Let me know in the GitHub page.
Today, I open-source Scenario System Design (SSD), a 3 step system design that allows an engineer to design a sustainable system architecture without spending a mind-boggling amount of time releasing the feature.
This all started during my days working in startups. I would be filled with ambition, designing a system so grand and powerful, able to cater to traffic in the millions. You see, startup founders would, naturally, fill me with dreams of a wildly successful business. People will flock to our business really fast once we launch the product. I must be ready for the riches to come.
What turned out was a product that never saw daylight for more than a year, and the startup ran out of money. Even if I wanted to launch the product, the product was so incomplete that even if I launched it, it barely even accomplish its primary function. It doesn't serve anybody.
And so, the startups fail. I wouldn't say that it's entirely my fault or I'm blameless. But this immature naive gullible me got a little too ambitious and tried to build a cathedral on pennies.
Then came jobs that required me to deliver results in record time.
Naturally, I would find the quickest and shortest way to accomplish the task. With such time pressure, there simply isn't any leeway to think long term. I was caught up in the moment, despite a tiny voice bugging, that my decisions today would haunt me sooner or later.
And true enough, as the technical debt pile on and on, it eventually led to a system so fragile that changing anything could easily break the application. It led to a point where I spent more time bugfixing than delivering features. I would going to business leaders telling them how the system needed a revamp. The leaders would be so frustrated at a system that frustrates the customers. I would be constantly fixing one bug after another while the competitors are delivering feature after feature. And you think that maybe if I adopted DevOps, it will help to speed up delivery? No No No... I simply incurred technical debt faster, launched faulty apps faster, and frustrated more audiences faster.
You see, DevOps is good at streamlining development to production, but it doesn't make up for poor design. If the application you designed cannot work without redesigning the database schema or refactoring large swathes code, no amount of DevOps will make it work. You must spend this manhours to revamp the application. In a fast-moving business landscape that requires us to be nimble, the heavy footsteps of a revamp ain't getting any greenlight.
So after all the frustrations I've gone through, I got thinking, what if we could prevent this situation in the first place? I wouldn't want a design that incurs all the technical debt, but I ain't going to make the mistake of designing a grand system that never sees the light of the day. It's through these many trial and error across many projects that I've honed this 3 step process I call Scenario System Design.
It starts with coming up with every conceivable scenario for the feature you are going to build, very similar to the startup days where I would go about building any sort of scenarios I was so hyped up about. The second step would involves designing a grand system that supports the scenarios you've conjured up in your brainstorming session. But the third and final step is what makes thing practical, the implementation step, which strips out anything in the grand design that you don't need to build now.
And thus, I teach SSD to my fellow peers in the companies I work for. It is through my interaction with my peers that I noticed not many engineers practise SSD. After all, the entire industry has pretty much standardise on 2 methodologies that doesn't encourage building in such manner, Waterfall and Agile methodology. While both of these are good on its own right, it doesn't solve the fundamental problem, that you need to build a system designed for tomorrow, but built for today's environment.
But the most important factor to why I'm sharing this out in the open today, is due to AI.
Before AI was doing much of engineer's day-to-day task, juniors in the company would spend much of their time writing codes, honing those skills towards writing efficient and maintainable codes. A good design wouldn't be at the top of their mind. A senior would typically step in to design the system and hand the blueprint down to the young padawans to implement. The seniors would also review the codes to see that they have written an efficient and maintainable codes, spotting silly mistakes like N+1 or major security lapse.
But AI has changed the entire engineering career ladder. AI is generating far more efficient and AI-maintained codes than our junior engineers are doing. That meant scope of work juniors handled is no longer relevant in the ladder. What're juniors suppose to do now? What do engineering managers do with the juniors in their team? Why would businesses bother hiring juniors? These were the questions I've been asking myself and my team lately. It makes no sense that their role is simply to prompt AI to perform work. Anybody can prompt now.
As such, I have started introducing SSD to the juniors in my team. What used to be in the domain of senior engineers and tech leads, now I bring my years of knowledge and experience down to juniors. And for the foreseeable future, this is what juniors will be trained upon. What will be the contributions of seniors and tech leads then? That I do not know. But one thing for sure, every engineering role has changed for good.
This reminds of me an old case study highlighted by MIT economist David Autor about bank tellers and Automated Teller Machines (ATM). Before the advent of ATMs, tellers were the frontliners handling every banking needs, from simple deposits and withdrawals to bank loan underwritings, they handle it all. But when ATMs started proliferating, every banker started worrying about their jobs. If the machines are taking over their job, what will the bankers do for a living? Many predicted the loss of many banking jobs, and sure enough, many roles in the banks were eliminated. But instead of the banking industry shrinking, it actually expanded. The bankers instead went on to take on much more complicated work, such as retirement planning, wealth planning, and so on. While the ATMs handled the bulk of mundane work, the tellers went on to handle higher-value tasks, which in turn increased the demand for banking services and the banks increased hiring to meet the demand.
Citation:
Are we seeing the same thing in engineering? I suspect so. And so, I give you SSD, to help transition teams towards higher-value contributions to their organisations they work for.
I've been watching the debate between those who view AI as an existential risks and those who dismissed it as fear-mongering. It's fascinating that the media reports 2 extreme views, but not much in the middle. That shouldn't come as a surprise anyway. Mediocre events don't generate enough buzz.
I was just looking at the report by Effective Altruism and it shows that AI repeatedly choose to harm humans to stop pain. Now, that is a very scary thought. Whatever happened to Asimov's 3 Laws of Robotics?
Then I read up about a new class of worm, Morris II, that targets email assistants. A self-replicated virus could easily spread like wildfire and the only known way to stop this threat is to disable AI assistant. It's like stopping a flu by getting rid of humans.
These events have strengthen my view that AI has the potential to be very harmful to humanity. Extinction? Maybe. But definitely has the potential to be harmful. It's not as easy as shutting down the AI lab if the lab cannot contain the danger. Time and again, we see AI models behaving in unexpected ways. Even some of the best AI labs have to physically cut off internet to its models for safety.
Let's take a simple test, see if you can spot the problem. You have a paperclip-making machine and you want to make as much profit from it as possible. One day, you decided to add AI to this machine and instructed it to "make as many paperclips as possible".
If you can't spot the issue, that's the danger. We humans are terrible at understanding dangers well.
Let's try another experiment to prove my point. Name me the most dangerous non-weapon, human invention. It's not nuclear bomb, since that's considered a weapon. If you say car, good guess, but no. It's the cigarette. That tiny stick that you lit up, kill 1 person every 3~4 seconds. This is the most ironic thing humans do. We take great strides to ensure air travel is safe, stay safe from shark, use sharp objects carefully, yet do almost nothing on the very thing that is manufactured to kill us. Yes, cigarette's are made to kill the consumer. Death is not the result of misuse, overdose, or accident. It is the statistical expectation of correct use.
So unless someone can objectively justify that AI poses no risk, I simply cannot accept an opinion that AI is guaranteed safe.
So Anthropic has released Claude Haiku 5.5. According to Artificial Analysis, it's intelligence score is 43. Not bad!
I'm using Qwen 3.8 Flash Next, which has a score of 40 locally. So I'm getting near Haiku intelligence at the cost of electricity.
But of course, there are trade offs. My Ryzen AI Max can only do inference speed between 4 to 22 tokens per second. The prefill speed is also really long. Now, on my wish list, I wish there's a Qwen 3.8 35B A3B, but American- or European-made. Better yet, Singapore-made.
It's been so long since I wrote a new piece. That's due to the hectic lifestyle I lived since I got married and my wonderful daughter is born. Juggling between fulfilling my work duties, being a loving husband and a doting father, it's been a struggle. Even my personal downtime is also compromised.
But recently, with the help of AI, I'm starting to reclaim back my time. It started with Google's Gemini Chat. To be accurate, Gemini wasn't the first LLM I used, but it's the first LLM that I use extensively for productivity. ChatGPT was the first, but it was more like exploratory usage.
Initially, Gemini was merely a tool to learn something really fast. Instead of scouring the internet with search engines to look for answers, Gemini could give me the answer within 1 minute. It used to be that I had to piece the information puzzle together one by one, page by page, site by site. What took hours was done in minutes.
Eventually, I started feeding it some PHP codes. It started to give me good solutions in tiny code snippets. It was readable and understandable. After all, when the code is 10 lines, it's easy to vet through line by line within 5 minutes. But then, that's the part I have a realization. While it took me 5 minutes to fully digest the code, Gemini took less than 1 minute to generate not just the code, but also the conversation explaining the code. This was the start of me handing over my over-20 years of engineering to AI.
Next came larger codebase. I started feeding entire files to Gemini to analyse and generate code. Since Gemini limited uploading 10 files at a time, I couldn't simply drag-and-drop. Some files were way too large. I could extract code snippets manually, but it turns out that the time I take extracting code snippets takes far longer than uploading 20 files, 10 files each time.
I also happen to have a deprecated Tabnine subscription, which was simply a VSCode extension and it was just plain old chat wrapper with some Claude models. Tabnine had restrictions, so I often hit rate-limits. But still, I made use of it as much as possible by combining Gemini and Tabnine's models intelligence by facilitating both of them to argue the best way to approach problems. Then came the day I anticipated. Tabnine cancelled my deprecated subscription and refunding any remaining unused amount.
At this point, I also started a new project and, this time, Gemini built the entire app without me writing a line of code. I was still using Gemini Chat up to this point. So it was a lot of drag-and-drop and copy-pasting between Visual Studio Code and Gemini Chat. Other administrative task include creating new files and folders, and performing git actions. Even git messages were written by Gemini at this point. But not a single line of code was written by me. It worked!
It is also around at this time that OpenClaw started to gain in popularity. After reading the horrors of security vulnerabilities, I decided to approach OpenClaw with extreme caution. Till today, I never expose OpenClaw to any chat app. I've also heard about the token burn using OpenClaw, so I used my company's Gemini subscription to experiment with OpenClaw. At that time, Gemini CLI had a generous limit and it could be hooked up with OpenClaw. This is the moment I realise the capabilities of Agentic AI. This is the moment everything changed.
Agentic AI is the future!
At least for the foreseeable future. So I poured by heart and soul towards this new shiny thing: Agentic AI. I was deeply amazed by Gemini's agentic capabilities riding on OpenClaw. This time round, instead of my usual drag-and-drop to Gemini, I simply pointed OpenClaw to my codebase and described my problem. Almost instantly, my problem was solved. It even ran tests to confirm that it works.
But of course, such generosity from Google couldn't last. Google tighten the rate limits to the extent that OpenClaw was unusable. I was almost always hitting rate limits. But I wasn't about to give up so soon. I knew I must eventually pay for my usage, but I wasn't sure how expensive it's going to be. Since I'm already an AWS customer and AWS bedrock host many models, I decided to try out Nova and Claude.
Both Nova and Claude is good. It solved every problem I threw at it, though it needed some guidance with context from time to time. I guess the problems I needed solving isn't that difficult, so both models worked fine. But strangely, when the first bill came, Nova cost more than Claude despite that Nova's token costs was lower than Claude. It turns out that OpenClaw didn't use cached prompts, so every prompt was charged at the input non-cached rates.
So I switched to using Claude exclusively. I used mostly Haiku 3.5, since most of my prompts ain't some complex math. Then the 2nd bill came. I was burning ~USD100 a day. Ouch!
That's it. I had to stop using hosted models. They are fast and good, but it burned a whole in my wallet. This time round, I focused on running local models. Since I already have a RTX 4060, why not make good use of it. Turns out, easier said than done.
Models that can fit inside the limited 8GB vRAM simply ain't smart enough. It can hold a basic conversation, but that's it. Getting it to power OpenClaw is nearly impossible. Offloading some of the weights to RAM just slows the model and makes the whole system unstable. Like the whole computer gets unstable. The best I could use my GPU for is to run Ornith1.5 9B to generate git messages through Zoo Code.
So my mission is the find a way to run larger models on dedicated hardware. The largest vRAM non-Apple, consumer-grade device I could get my hands on at that time was Ryzen AI Max+ 395, costing me more than SGD4000. But hey, if I was burning USD100 a day running Claude Haiku, I would easily break even within 2 months, running to smartest open-weight model I could fit.
And there I got it. The day it arrived, I was so eager to try it. I allocated the maximum amount of RAM I could to GPU and get my hands dirty with the first LLM. Since it was my first attempt, I just wanted it working as soon as possible. Ollama was my immediate choice. It was simple and effective. Most importantly, it simply worked.
But as smooth as it was with deployment, the price of this simplicity was the prefill and inference speed. Ollama comes with a tax on every prompt. After spending this money, I wasn't about to leave any performance on the table. So I figured out how to run Llama.cpp raw. No more middleman.
Turns out to be harder, but once I get the execution commands right, I just needed a script to execute the same way every time.
These days, I'm running Qwen3.8-flash-next. It's a bit on the slow side, but it works as good as Claude Sonnet 5. I'm contemplating whether to switch to Ornith1.5 35B A3B. It is suppose to be slightly worse than Qwen3.8-flash-next, but runs at far better speeds. But for now, as long as it works and gets things right, maybe I'll stick to Qwen3.8-flash-next for the time being.
I'm also using Cursor for my development. I've got to say, the auto+composer token limit sure is generous. Of my months of subscription since May, I only ever consumed 100% of my monthly allocation once. I could use my Qwen3.8-flash-next with Cursor, but I don't see the need to. Why use my local model when I still have plenty of tokens left with Cursor?
My electricity bill has certainly gone up. I'm using at least an extra 100kWh every month, with amounts to more than SGD34 every month. But this amount is far easier to swallow than the USD100/day I was spending before. I would happily pay my utilities over AWS bedrock. The other effects of running my local model is that the GPU is heating up my room, so I am using a bit more air-conditioning than I usually would. Still, happy to pay this bill over AWS bedrock anytime.
I don't know when I can write my next post, but I guess this post is a sign that I have managed to reclaim back a little bit of my time so far. I hope I can continue to reclaim further more. I won't let AI write my posts. After all, you're here because you want to hear my voice, not some version of AI I control. But the images will be AI generated.
Greetings Earthlings , Shurn the Awesomer is here to give you an awesome time.
This little site is a record of my life, opinions, and views. I'm mainly writing about Technology & Gadgets, Busting Creationist Myths, and other philosophical stuff.
This site is done using CakePHP.
With this uptime, how much more can I be proud of to showcase to the world? This uptime monitoring is brought to you by StatusCake since 13th May 2017.
I will always check for copyright usage before using any materials on my site. Whenever due, credit shall be given.
However, if you notice that I may have infringed on any copyright material. Please do not hesitate to contact me. All works of every artist deserves to be honoured and respected.