Hey, it's Bernie!

What would a truly localized Singaporean LLM look like? Beyond just Singlish, I explore how a 'fellow citizen' AI could navigate our unique cultural nuances.

---

As I'm going through the free Stanford Online course CME295, I thought about Natural Language Processing (NLP) and how we currently tokenize English text to feed into LLMs. If we transform words into numbers, how do we add in culture nuances, slangs, etc that are unique to a country so that we can create a sort of naturalized citizen? What is Singapore currently doing?

So I asked Gemini for some thoughts: so if I am a citizen of a country who wants to help build a realistic LLM for my country where we have 4 languages and we want to not only create a balanced LLM, but also to incorporate our country's culture and slangs into it so that the LLM performs like a "fellow citizen" and knows all our nuances and history. How should I start?

Gemini's answer: “This is a noble and technically fascinating mission. Building a “Sovereign LLM” is the move from a model that knows about your country to a model that is of your country.

To make an AI feel like a “fellow citizen,” you have to move beyond just scraping the web. You need to follow a four-stage blueprint:

• Collection: Gathering multicultural datasets.
• Curation: Cleaning data to avoid “Data Poisoning.”
• Alignment: Teaching the AI nuances and cultural slangs.
• Evaluation: Testing if the AI actually sounds like a “Fellow Citizen.”

...

Be careful with “Data Poisoning.” If you only scrape one political side or one ethnic group's slang, your “Fellow Citizen” AI might end up being biased or exclusionary. Balance is your hardest job.”

It then went on to suggest some open-source models to start with and actionable next steps.

[NOTE] I'm using this blog article to document the start of my journey. It will be a multi-year quest of mine. I know Singapore is already embarking on building her own LLM or LLMs, but wouldn't it be great if I can create my very own Singaporean friend, a super knowledgeable, kind and caring one at that.

Cool idea?

#ai #LLM #NLP #SovereignAI #Singapore #Singlish #AISingapore #CME295 #LearningInPublic

Today I submitted my technical assessment for the AIAP – aka AI Apprenticeship Program by AI Singapore. It was like I went into a cave and vibe-coded with Claude for 6 days. The ball's in their court now.

This morning I felt a sense of emptiness.

There has been some collateral damage as my Bug Farm was not tended to diligently, causing the mealworm lifecycles to overlap in the same bins. If you don't know what I'm talking about, it's ok.

So, today I'm going to spend most of the day tending to my worms, then for Claude I'm going to test out some stuff:-

• Learning more about skills and structuring my coding workflows to utilize this feature
• Changing Claude's setting to “brutally honest”
• I have a mini project I codenamed “Under the Hood”, which I will start on and will update you soon

Oh yes, editing my video for Bernie Studio Delivery

Have a great week ahead!

#reflection #PositiveThinking

Can a solo-builder contribute to Singapore’s national AI mission? I’m pursuing an AI Singapore apprenticeship to help transform our local talent into a global export.

---

I recently had a realization while watching Dispatches from Elsewhere: staying quiet is a form of selfishness. In the show, Fredwynn tells Peter that by not sharing his perspective, he’s withholding something valuable from the group. That hit home. For too long, I’ve been a quiet solo-builder, but as I pursue an apprenticeship with AI Singapore (AISG), I’ve realized that my mission, and the knowledge I gain, belongs to the community.

Singapore has no natural resources; we only have our people. I believe AI is our unique opportunity to transcend our physical limits and export local talent to the world. To get there, I’m attacking this from three angles: building real-world solutions, earning technical certifications, and finally, refusing to be 'selfish' by sharing every step of this journey in public.

• First I will just solo-build, or team-build (I don't have many friends to start with actually), help my friends solve their issues, learning and building at the same time.
• Secondly, I will build my credibility by getting certifications, skills and knowledge
• And third, I will resolve to share more about my journey, because I recently realized that if I don't share, I'm being selfish with my knowledge. So ... I will share more, here on my blog, and also look out for upcoming videos on YouTube.

If I don't get the apprenticeship I will still keep building.

Thanks for reading.

#reflection #ai #apprenticeship #AISingapore #AISG #SingaporeTech #SoloBuilder #AIGovernance #SmartNation #CareerPivot #LearningInPublic

#music #playlist #ai

Nobody: What did the server say? Me: I got a feeling, somebody's pinging me.

#lame

After two months of diving into machine learning, I’ve realized that data science is like performing an autopsy on a business problem to find the clues hidden in the facts. By treating data like a detective mystery, we can stop guessing and start building solutions based on real-world evidence and predictive confidence.

At the start of 2026, I jumped head first into the world machine learning (ML) and data science (DS), and I realized that I'm starting to look at the world with a new lens, and add-on if you will, where every problem starts with data.

I know, some of you reading this might say “no Bernie, we always start with identifying the business problem, and ask ourselves what are we trying to solve.” Yes, that too, but after that, you would hire a data scientist right? That's where I'm starting my story.

Why Data Science is like a Medical Drama

The clues are in the data. With every business problem, the data is like a dead body on an autopsy table, ready for the data scientist to slice and dice, looking for clues as to why there's a problem in the first place. I've always loved a good detective and/or medical drama, and now I know why. ML and DS is also like that. Find the clues, solve the mystery. And after we solve the mystery we can start building a solution based on facts, which can then point us into the right direction to collect, not just more data, but the RIGHT data. Ultimately we would like to be able to predict with higher confidence, the consequences of our future decisions and actions, so we don't make the same mistakes that would cause said problems to arise again.

So many examples from the real-world, like

• Hotel Industry: Reducing “no-show” rates by identifying patterns in booking cancellations.
• FinTech: Detecting fraudulent transactions before they clear by spotting anomalies in spending.
• Education: Analyzing student performance data to intervene before grades drop.
• Marketing: Auditing marketing spend to see exactly where the ROI is failing and why.

It can help us create systems and processes for better business ROI.

I've always wondered about such things. Why are things the way they are. Almost everything we see around us is a consequence of someone's decision and action – what made that person come to that choice? And by understanding their motivations (through the data) we can make better decisions for a future that we want, whether good or bad.

With the right and big enough dataset, we can predict (and solve) almost anything.

#ai #data #reflection #MachineLearning #DataScience #BusinessROI

Now we code.

[UPDATE] I did a classic noob move to use Opus to help me tweak a css file. Session ended pretty fast. Note to self: Just use Sonnet.

#claude #ai

Deliveroo has exited the Singapore market. It impacts me directly because I've been a Deliveroo rider for the last three plus years – it was my side hustle as I sorted my shit out and figured out what I wanted to pivot to. Even though I'm also on Foodpanda, another food delivery platform, Deliveroo is the best as it has a more direct fee system, for example, it gives you fee boosts vs quests, and the ONE THING that makes it so much better is that it allows riders to “reject jobs without penalty”.

Anyway, the announcement about the exit gave little runway for us all in Singapore.

Announcement made on Feb 26, and exit on Mar 3. Read the article here

WOW. And Ouch.

I'm not even sure the execs at Door Dash (the guys who bought Deliveroo) even know where Singapore is. Anyway, to them we are just a non-performer on a spreadsheet, bound for deletion to make the “books” look better.

I decided to make a video where I travel to almost every zone in Singapore in the little time left they have here in our country, as a symbolic gesture to the platform that allowed me to earn some lunch money while I noodle away at life. I will link to the video here once it is done, so stay tuned.

#deliveroo #fooddelivery #reflection

#music #bts

As a self-proclaimed 'financial illiterate,' I used Gemini 3 to build a Python dashboard that tracks the Rule of 40, Altman-Z, and other metrics to find high-growth companies.

---

I am no good with money and have never really invested in anything other than one piece of property in my life. But I was thinking, is there a way to identify high-growth companies and be able to predict their ability to stay as leaders and invest in them. What would be the benchmarks and metrics to measure their worth on?

So I wanted to try and create a dashboard, which will tell me at a glance a sort of a health-check, whether a particular company is doing well or not.

I chose Gemini 3's thinking model for this. Apparently it “solves complex problems”. But my problem isn't really complex, I just wanted a python script calling an API and visualization using Streamlit.

Here's my starting prompt what are the key indicators of companies when evaluating their business growth potential and revenue model, and what kind of competitive analysis data should I collect to compare similar companies to separate the winners from the losers

What followed was about 2-3 hours of Gemini and I coding out a script which connects to yfinance to get real-time data, but now I heard that it isn't that reliable. Here's version 1 of the dashboard.

Project Specs:

• Stack: Python, Streamlit, yfinance API
• AI Collaborator: Gemini 3 (Thinking Model)
• Build Time: 2 days (from concept to prototype)
• Goal: Demystifying complex financial health checks for non-experts.

Check it out —> Bernie's Finance Dashboard

What it does

1. You pick a ticker you want to analyze
2. The dashboard will show you some meaningful financial metrics
   • The Rule of 40: Balances growth vs. profitability—a must-have for SaaS companies.
   • S&M Efficiency Ratio: Measures how much revenue you get for every dollar spent on Sales and Marketing.
   • Altman-Z Score: A “crystal ball” formula used to predict the likelihood of a company going bankrupt within two years.
   • Cash Runway: Tells you exactly how many months a company can survive if it doesn't make another cent.
3. There's also a way for you to enter a benchmark ticker to compare with your target ticker
4. If you are unsure of the what the metrics mean, there's also a quick guide on what those metrics measure

On to my next project! Enjoy!

#finance #ai #built #Python #Streamlit #FinTech #Gemini3 #SoloBuilder #DataVisualization #FinancialLiteracy #GrowthMetrics