Data Science Teacher Brandyn

01/05/2026

Congratulations to the Winning Team Yabba Dabba Data! At last nights AWS Jam.

That was an awesome experience and I was so fortunate to have an amazing team that together we used fixed AWS Bedrock Ai Agentic systems and we ended up winning by a mile :).

Thanks Hussain, Priya and Shivam. You guys were an awesome and really helped us win by a large margin. :)

Thanks Team :)

Its too bad we were working so hard that we forget to get pizza! lol

28/04/2026

We joke about AI taking over the world, but while we were laughing, 5 AI agents quietly unpacked their bags and moved into your daily routine.
It feels exactly like that meme where you're standing alone in an empty bathroom, and a guy walks in just to pick the urinal right next to yours.
AI has gotten uncomfortably close to our personal lives, and most of us didn't even notice.
I’ve been studying for the AWS Generative AI Developer exam lately, and let me tell you, this stuff is hard, and it’s okay if it doesn’t click yet.
But as aspiring data scientists, it is our job to peek behind the curtain and spot the invisible code running our day.
Here are 5 areas where AI agents have quietly taken over, and the hidden data science making it happen:
1. The Inbox Bouncer
Your email isn't just a static spam filter anymore; it's an active NLP agent.
(NLP stands for Natural Language Processing, which is simply a way we teach computers to read, understand, and categorize human words).
This agent reads the context of your emails and silently decides what is important enough for your primary inbox based on your habits.
2. The Fraud Detective
When you buy a coffee, an agent runs an Anomaly Detection model in milliseconds to make sure it's actually you.
(Anomaly Detection is just a math system trained to ring an alarm when data looks out of character for your normal daily routine).
3. The Rideshare Negotiator
Ever notice Uber or Lyft prices jump while you're just staring at the screen?
That’s an agent using Reinforcement Learning.
(Reinforcement Learning trains an AI kind of like a puppy, giving it a digital "treat" or reward every time it makes a highly profitable choice based on live city demand).
4. The Social Media DJ
TikTok and Instagram don't just show you a list of popular videos; they use a Multi-Armed Bandit algorithm.
(This is a casino-inspired math trick where the app constantly tests new, unseen content on you while safely showing what you already like, just to figure out exactly what keeps you hooked).
5. The Texting Mind-Reader
Autocorrect is dead; your phone now uses lightweight LLMs to predict your next thought.
(LLMs, or Large Language Models, are massive prediction engines trained on the whole internet to logically guess the very next word in a sentence).
I remember the first time I tried to code a basic text-prediction model, and my "AI" just repeatedly spit out the word "banana" over and over again.
I felt so defeated, and I questioned if I was cut out for this field.
So if you are reading about Generative AI today and feeling a little overwhelmed, please be kind to yourself.
Take your time with this.
Let’s look at it one more time, break it down, and remember that every expert was once a beginner staring confused at a broken script.
You have the immense potential to not just use these invisible agents, but to build the ones that shape the future.

25/04/2026

Stop trying to build the perfect AI application on your first try. Do this weirdly simple step instead.
I remember my early days trying to build machine learning models, thinking I had to write the most complex code ever on day one.

It was a complete disaster, and I spent hours staring at a broken screen with no idea where I went wrong.

(This stuff is hard, and it’s okay if it doesn’t click yet!)

Here is the weirdly simple step: start with an "ugly baseline."

A baseline is just a simple, basic version of your project that you use as a starting score to compare all your future, complex versions against.

I always tell my data science students this exact same rule whether they are cleaning data, doing an analysis, or building a neural network.

A neural network is a type of machine learning model inspired by the human brain that learns to recognize complex patterns.

I recently had to swallow my own medicine while building "AI Teacher Brandyn," my new educational AI.

I am building a Multi-Agent System for it.

A Multi-Agent System is just a fancy way of saying a team of specialized AI bots working together, instead of one giant bot trying to do everything.

Right now, my architecture is intentionally basic with just two main bots: a "Learn Agent" to guide students through the curriculum, and a "Test Agent" to give practice quizzes.

Overseeing them is a "Routing Supervisor Agent," which acts like a traffic cop, reading the student's prompt and sending it to the right bot.

It is so tempting to add ten more agents to do a million cool things, but here is the rewarding secret: you have to establish that rock-solid baseline first.

I am taking my time experimenting with and fine-tuning these two simple agents.

Then, I validate their performance using a testing technique called "LLM-as-a-judge."

LLM-as-a-judge is when you use a very smart AI model to grade the homework of your AI system, scoring how accurate and helpful the responses actually are.

Because I have a baseline score for a simple setup, whenever I tweak the architecture or add a new agent later, I have something to compare it to.

This guarantees I am creating a real-world improvement in the output, rather than just adding complexity that breaks the system.
Take your time with this. Let’s look at this one more time: build simple, test, validate, and then add complexity.

You have the potential to build incredibly powerful and advanced tools, but remember that every great data project starts by mastering the simple steps first!

PS: I once spent an entire weekend trying to fix a deep learning model, only to realize I forgot to normalize my data... a painfully simple step I skipped because I was rushing to do the complex stuff!

21/04/2026

The difference between an AI Agent and Agentic AI is the difference between a microwave and a personal chef.
Most people are still treating AI like a fancy Google search.
They think an AI Agent and Agentic AI are the same thing.
I used to think so too, until I tried to automate a data pipeline and realized my "agent" was just sitting there waiting for me to hit "Enter" like a lazy intern.
This stuff is hard, and it’s totally okay if it doesn’t click yet—most of the industry is still arguing over the definitions!
Let’s look at this one more time using the image I attached.
An AI Agent is like a smart thermostat.
AI Agent (Task-Focused): A software entity that can observe its environment and take a specific action to reach a single goal.
You tell it "Set the temperature to 21°C," and it does it.
It’s great at ex*****on, but it has a narrow "blindfold" on.
It won’t care if the window is open or if electricity prices just tripled—it just does the one job you gave it.
Agentic AI (Outcome-Focused): A system of multiple agents and tools working together to own a "big picture" result without you holding its hand.
Think of the right side of the image.
The system doesn't just "set the temp."
It checks the weather forecast (Data Retrieval).
It looks at your calendar to see when you're home (User Context).
It monitors energy prices to save you money (Optimization).
It even talks to your appliances to balance the load (Orchestration).
Agentic AI doesn't just do the task; it handles the workflow.
I remember the first time I tried to build a "Simple" scraper... I built an agent to get the data, but I had to manually clean it, manually save it, and manually email the report.
I was the "orchestrator" doing all the hard work!
A truly Agentic System would have seen the dirty data, spun up a "cleaning agent," and realized the report needed to go out every Friday without me saying a word.
Here is the hidden gem for your next interview:
The "secret sauce" of Agentic AI isn't just the LLM—it’s Reasoning and Tool Use.
In 2026, the best Data Scientists aren't just writing Python; they are designing Orchestration Layers.
This is the "brain" that decides which tool to call and when to "loop" back if the first result wasn't good enough.
Take your time with this.
Moving from "Prompting" to "Designing Systems" is a big jump.
But once you start thinking in outcomes instead of tasks, your value as a developer skyrockets.
You’ve got the brain for this—keep building.

18/04/2026

The hardest skill in data science today isn't writing Python anymore—it's reading it.
Everyone thinks large language models, or LLMs, replaced the need to learn the basics.
(An LLM is just the fancy text-predicting AI powering tools like ChatGPT and Gemini).
But from continually doing projects with my students, I'm seeing the exact opposite.
Yes, co-writing code with Gemini in Google Colab gets us to the finish line much faster.
It handles the boring, common parts of a project like a champion.
But I remember the first time I let an AI write a complex data cleaning script for me without checking it closely.
It completely dropped half my dataset because it misunderstood my logic!
This stuff is hard, and it’s okay if navigating AI code doesn’t quite click yet.
Here is the hidden trap most beginners fall into.
As soon as you inject your own ideas and your project becomes a unique research concept, the AI loses its high-level accuracy.
It starts making assumptions.
What ends up happening is you have to read 15 to 20 lines of AI-generated code just to spot the 2 or 3 sneaky mistakes.
Or you have to catch exactly where it started branching in a completely wrong direction for what you planned next.
So even though we physically type way less code these days, the demand on understanding what each line actually does has skyrocketed.
You have to be an editor now, not just a writer.
Let’s look at this one more time with a quick, rewarding tip you can use today to fix this AI hallucination issue.
When prompting your AI for a data transformation (which just means mathematically altering your data, like shrinking extreme outliers so they don't skew your results)...
Don't just ask it to "clean the data."
Instead, explicitly list out the exact columns you want to transform and the specific method you want for each.
If you tell it exactly which columns need a log transformation and which need outlier capping, it understands your intent perfectly.
It will rapidly generate the slightly varied, customized code you need for each individual column without guessing.
Take your time learning these fundamentals.
It's not about memorizing syntax anymore; it's about knowing enough to steer the ship.
If you want to learn the basics in a simple way to accelerate your whole learning process, check out our free path at DataSimple.
https://www.datasimple.education/datasimple-data-learning/python-data-basics-free-learning
You have incredible potential to build amazing things once you master these foundations, so stick with it!
PS: I still look up how to drop a column in Pandas sometimes, so don't ever sweat the small stuff or feel like you need to memorize everything!

15/04/2026

I finished a full-time university degree, many AWS Certications all while working a 40-hour week and kept my marriage happy—without "grinding."
Check out the full conversation on Mike Vardy's Podcast a productive conversation: https://www.youtube.com/watch?v=0Lg0nOpVSjg.
Most people think productivity is about squeezing more juice out of a lemon.
But if you’re a "Forager" like me—someone with a restless, high-energy mind—traditional focus techniques feel like holding your breath underwater.
I remember sitting on my apartment floor years ago, stuffing 500 envelopes a week for a financial job I hated.
My hands were moving at 5 mph, but my brain was spinning at 100 mph, and that friction made me want to crawl out of my skin.
I wasn't "broken" or "distracted." I was under-utilized.
I realized that if I gave the "wandering" 15% of my brain a simple task—like a narrative movie—the other 85% could actually do the rote work in a "Mechanical Trance".
This is what I call Parallel Processing: layering a low-load task (like laundry or commuting) with a high-load one (like learning a new language).
The secret "hidden gem"? It’s called Cross-Modal Stacking.
Your brain has separate "pools" for different senses. You can’t write an email and listen to a podcast at the same time because they both fight for your Language Center.
But you can listen to a lecture while walking or cooking because those tasks use different neural "lanes".
I used these exact methods to knock out AWS certifications and a degree while working full-time, all without missing dinner with my wife.
You don't need more willpower; you need a better internal "operating manual" for the 21st century.
Take your time with this—it's okay if it doesn't click immediately, but know that your restless mind is a high-performance engine, not a flaw.
You have the power to be an architect of your own focus, not just a reactive participant in a loud world.

PS: Here is a quick win: Next time you have a "dead" 5-minute gap, don't doom-scroll. Use a "Strategic Switch" to knock out one tiny, pre-planned task from your list. It’s the dopamine hit your brain actually craves.

Much of the modern productivity conversation revolves around a single ideal: deep work. The ability to focus intensely on one task at a time is often present...

14/04/2026

My friend tried to break my AI, and it taught me a $15,000 lesson in "cheap" tech.
My friend tried to break my AI today.
https://www.datasimple.education/ai-tutor-brandyn
He went full "troll mode" on Teacher Brandyn at DataSimple.education.
He wasn't looking for Python tips or Seaborn tutorials.
He was stress-testing the soul of the machine.
And it was the best thing that could have happened for my product.
I’ve been debating a technical choice lately:
Should I use Claude Haiku or Claude Sonnet for my "orchestration agent"?
An orchestration agent is basically the "brainy receptionist" that decides which specialist AI should handle your specific question.
Haiku is fast and dirt cheap.
Sonnet is the premium, heavy-lifting model with deeper reasoning.
I thought, "Hey, for just routing a message, Haiku is plenty smart, right?"
Wrong.
Watching my friend’s "troll" prompts, I realized something critical.
The first contact isn’t just about classification—it’s about character.
Haiku might have just tagged the message as "noise."
But Sonnet? It handled the trolling like a champ, stayed in character, and gently nudged the "student" back to data science.
It maintained the "Humble Guide" vibe I worked so hard to build.
In data science—and in business—we often try to optimize for cost too early.
But if your "first contact" layer feels like a rigid robot, you lose the user before they even see your best work.
This stuff is hard, and building these systems is even harder.
It's okay to spend a little more on the "brain" if it means your users feel heard, even when they’re just messing around.
Don’t cheap out on the bridge between your tech and your humans.
You’ve got the potential to build something amazing—just make sure it has the "wisdom" to handle the bumps along the way.

07/04/2026

Building an AI agent is fun, until it suddenly forgets who it is and starts talking like a generic textbook.
I remember the first time I set up a multi-agent tutor for DataSimple.
Ai Teacher Brandyn
https://www.datasimple.education/ai-tutor-brandyn
A multi-agent system is just a setup where different AI programs work together, like a manager delegating tasks to specialists.
I had carefully crafted a warm, encouraging "Teacher Brandyn" persona for the AI to use with students.
But during testing, it completely broke character and spat out a dry, robotic summary of my database.
This stuff is hard, and it’s okay if cloud architecture doesn’t click for you immediately.
I spent hours pulling my hair out before I realized AWS Bedrock was actually trying to "help" me.
To save compute time, some platforms will completely bypass your carefully written instructions if they think your agent is just a simple data-retrieval bot.
The hidden gem to fix this?
You have to manually enable "User Input" in the settings to force the AI's orchestration brain to wake up and read your persona instructions every single time.
But once I fixed that, I hit another wall: my AI examiner gave a 3-question quiz, but then confidently asked "Question 4" anyway.
It got stuck in an autoregressive loop.
An autoregressive loop is when an AI just keeps generating text based on the pattern it just wrote, instead of knowing when to stop.
Large Language Models are notoriously bad at counting without explicit tracking.
The forensic insight here is to give the AI a "scratchpad" using XML tags.
XML tags are simple text markers used to structure hidden information, like .
Force the AI to silently write The user answered question 3, I need to stop the test before it generates its final response to the user.
The user interface strips out the tags so the student never sees it, and the AI stays perfectly on track.
Take your time with this, because engineering an AI's logic flow takes a lot of trial and error.
But once you master how to guide an AI's internal reasoning, you become an incredibly dangerous problem-solver in the modern data landscape!
PS: Always test your AI prompts by asking it to do something slightly outside its instructions—like replying with just a single letter "B"—it's the absolute fastest way to find exactly where your guardrails break.

01/04/2026

Learning to code isn't the hardest part; the real nightmare is figuring out exactly what project to build next.
Stop guessing what data skill you need to learn today—I just built an AI tutor to tell you.
Ai Teacher Brandyn
https://www.datasimple.education/ai-tutor-brandyn
I am beyond excited to announce the release of AI Teacher Brandyn to help you learn how to code!
The transition from a traditional computer science mentality to a modern, data-driven workflow represents a significant psychological and technical hurdle for many learners entering the analytics space.
This stuff is hard, and it’s completely okay if it doesn’t click right away.
I remember the first time I tried to learn programming; I felt completely overwhelmed and lost in the weeds.
Historically, programming has been taught through a trial-by-fire methodology, focusing heavily on raw functions, loops, and abstract algorithms in isolation before ever introducing practical applications.
The curriculum philosophy at DataSimple.education directly challenges this antiquated paradigm.
The platform operates on the fundamental belief that learning data science with Python can be highly accessible, provided the student is guided through the difficult mental transition to a "coder mindset".
That is exactly why I built this tool for you.
Phase 1 of development is officially complete, getting the learn function active to assist students directly with the lessons available on DataSimple.
If you are not sure where to start, this tool is designed to evaluate your level and guide you to the most valuable lesson.
It can help you navigate the deep pedagogical knowledge from nearly 100 Guided Projects, Data Tips, and YouTube video lessons.
This isn't just a generic chatbot; it is an advanced Retrieval-Augmented Generation (RAG) system.
RAG simply means the AI actively looks up precise facts and curriculum materials from a trusted database before it answers you, rather than just guessing or making things up.
This AI tutor was explicitly designed to mimic the pedagogical approach of a supportive data science instructor.
It goes way beyond answering generic coding questions by actively retrieving precise curriculum materials, guided projects, and methodologies taught directly within the DataSimple ecosystem.
Here is a rewarding hidden gem about how it works under the hood.
We strictly isolated the database as a "Ground Truth" repository, meaning it only embeds complete, 100% correct, runnable solution notebooks.
But to prevent the AI from acting as a simple code dictionary and giving away these solutions to struggling students without effort, we gave it a Master Prompt.
The Master Prompt instructed the AI to act as a dedicated tutor, utilizing the retrieved ground-truth projects to understand the context, but requiring it to guide the student to the answer using iterative hints.
Take your time with this tool and let it guide you step-by-step.
Let’s look at your educational roadmap one more time, together.
You have incredible potential to master these skills and seamlessly transition into the new modern data world.
PS: I almost threw in the towel building this when my data extraction pipeline kept crashing due to messy, hidden HTML tags in the raw video subtitle files.
Sometimes the messiest data teaches you the most valuable engineering lessons!

24/03/2026

Management thinks AI is a magic wand, but it’s actually a high-pressure power washer.
That meme is painfully accurate.
We’ve all seen it: management hoping a new "AI Agent" will magically erase a decade of messy code and "spaghetti" data.
Spaghetti Data: When your data sources are so tangled and poorly documented that moving one piece breaks three others you didn't even know existed.
I remember back when I started teaching, everyone said Scikit-learn was the final frontier.
Then it was TensorFlow and Neural Networks.
Now? It’s AI Agents.
I’ll be honest—even as an "expert," I wake up most mornings feeling like a beginner because the finish line keeps moving.
But here is the forensic truth: AI isn't a delete button for technical debt.
In fact, it's often a magnifying glass for it.
5 Reasons AI Won't "Just Fix It"
1. The "Garbage In, Agent Out" Loop
AI agents rely on the data they can access. If your underlying data is a mess, the agent will just automate the process of making bad decisions—only much faster than a human ever could.
2. The Hidden Maintenance Tax
Every AI model you deploy is a living thing. It requires "Model Drift" monitoring.
Model Drift: When an AI's performance drops over time because the real world changed but the data it was trained on stayed the same.
3. Integration Complexity
Adding AI to a broken system is like putting a Ferrari engine into a rusty 1990s sedan. The frame (your infrastructure) wasn't built for that kind of torque, and things are going to snap.
4. The "Black Box" Trust Gap
When an AI makes a mistake in a regulated industry, you can't just say "the robot did it." You need "Explainability."
Explainability: The ability to look at an AI’s output and trace back exactly which logic or data points led to that specific result.
5. The Human-in-the-Loop Burden
AI doesn't eliminate work; it changes the nature of it. Instead of doing the task, you are now auditing the task. This requires a higher level of skill and constant upskilling.

18/03/2026

The "hidden" reason your computer vision model is failing (and how to fix it).
I spent $300 of my own money and ran 36 separate experiments just to realize I was asking my AI to do the impossible.
I was working on my thesis, trying to predict if an Amazon product would be returned based on the "honesty" of the image compared to the description.
If the photo is a lie, the customer sends it back. Simple, right?
Wrong.
My model kept failing, and I couldn't figure out why until I used SHAP (SHapley Additive exPlanations).
SHAP is a tool that acts like a forensic investigator, showing you exactly which pixels or features your model is actually looking at to make a decision.
It turns out, for "Shoes," the actual shoe only took up 10% of the photo. The rest was just empty white space.
For "Phones," the model was staring at the gap between the front and back view of the device, rather than the device itself.
My data quality was poor because the "signal" (the product) was drowned out by the "noise" (the background).
I realized too late that I needed Bounding Boxes—which are essentially digital "fences" you draw around an object to tell the AI, "Hey, only look inside here."
This stuff is hard. I felt like a failure at experiment #20, and by #36, I wanted to quit.
It’s okay if your first (or thirtieth) attempt doesn't work. The breakthrough usually happens right after the frustration.
Don’t be like me and wait until the end of your project to audit your data.

5 Tips to Double-Check Your Data Quality Today
1. The "Human Eye" Test: Randomly sample 100 images or rows. If you can’t tell what’s happening in 2 seconds, your model won't either.

2. Check the Aspect Ratio: Ensure your images aren't being squashed or stretched during resizing, which turns a "Smartphone" into a "Pancake."

3. Audit the "Noise": Look for excessive white space or backgrounds. If the background is always the same, the model might learn the "room" instead of the "product."

4. Use Explainability Tools Early: Run SHAP or Lime on your first 5 experiments. Don't wait for experiment 36 to see what the AI is "seeing."

5. Class Imbalance Check: Ensure you don't have 1,000 shirts and only 5 pairs of shoes, or the model will just start guessing "shirt" for everything.

Check out the full breakdown of my struggle (and the final code) on my GitHub:
https://github.com/Brandyn-Ewanek/multimodal-ecommerce-return-prediction
You have the talent to solve these problems; you just need the right data to back you up. Keep going.

PS: I once spent three hours debugging a "broken" script only to realize I hadn't plugged my laptop in and it had gone into power-saver mode. We all have those days! Focus on the progress, not the perfection.