I love to teach. I post about #dataengineering and #mentalhealth daily. 

Many years of data engineering experience doing growth analytics at Facebook, security infrastructure at Netflix, and pricing and availability infrastructure at Airbnb. 

You can find me by looking up EcZachly on all social networks 

Join my Data Learning Platform https://www.dataexpert.io

Zach Wilson

Databricks is going to win “data” the same way that Meta won “social media” 

They are uniquely positioned because of the convergence of a few trends:

- the death of the data analyst
We are moving away from analytical architectures and towards action-oriented architectures. Instead of “the data engineer does ETL so analysts can look at it” it will be “the data engineer creates context for AI agents to do something” Actions will reduce timelines and speed up business velocity. Humans will no longer be the analytical bottleneck. 

- the crushing of data siloes 
Open table formats like delta and iceberg will make business able to do things that were previously though impossible! We’ll finally have all the context for our AI in one place. Entire business context + Agents = a waterfall of business value! 


I’m launching a paid Databricks boot camp on February 9th. Send me a DM with the word Databricks and I’ll get you the details!

Founder @ DataExpert.io | ADHD | 500k+ followers | Angel Investor

No alternative text description for this image

The first speaker lineup for the analytics engineering boot camp is going to be so amazing! 

These sessions are what truly separates my boot camp from “just another data course”

Data modeling is far from dead! It’s actually more relevant than ever. 

There’s been an interesting shift in the seas with AI. Some people saying we don’t need to do facts and dimensions anymore. This is a wild take because product analytics don’t suddenly disappear because LLM has arrived.

It seems like to me that multi-modal LLM is bringing together the three types of data:

- structured

- semi-structured

- unstructured

Dimensional modeling is still very relevant but will need to be augmented to include semi-structured outputs from the parsing of text and image data.

The necessity for complex types like VARIANT and STRUCT seems to be rising. Which is increasing the need for data modeling not decreasing it.

It feels like some company leaders now believe you can just point an LLM at a Kafka queue and have a perfect data warehouse which is still SO far from the actual reality of where data engineering sits today

Am I missing something or is the hype train just really loud right now?

I'm doing a live with Karan Dwivedi to talk about a cybersecurity boot camp that is launching on January 9th! 

This boot camp has an amazing capstone project where you do all of these on a company server!
- threat modeling
- detection engineering
- preventative controls 

Don't miss this webinar tomorrow!

January 9th Cybersecurity Boot Camp Webinar

Nirmal is a Data Scientist who yearns for solving problems that actually impact product strategy and business outcomes. As a self-described realist, he believes that with pragmatism and high-quality data, Data Science can be a very powerful tool. 

If you are preparing for ML interview, I have something to share: https://onlyoneoutlier.gumroad.com/l/decodingML 

Nirmal has experience working for both government and private industries. He had worked as Security Data Analyst for Department of Navy, and had also served in US Army. He loves working in the rare intersection between security and data science, where there are ample challenges to tackle.

If you have stalked me this far, you have to hit that follow button if you have not done yet :) Jokes a side, I love to share frequent posts on data science, machine learning, mentoring and career coaching, so following me won’t hurt.

Want to have some data science career chat regarding interviews, resume reviews or anything?  You're welcome to reserve a time at: https://topmate.io/nirmal_budhathoki

I also write data science blogs in substack, please subscribe at: https://onlyoneoutlier.substack.com/

Nirmal Budhathoki

It hurts when your girlfriend betrays you …

Hurts even more when p-value betrays you . 

When p is low, you don’t always glow 
A true Data Scientist will rather take it slow. 

A silly rhyme, p-value of you smiling is low,  but hey low isn’t the better 😆 

In interviews and real projects, I often hear:

“The p-value is < 0.05, so the feature is always important.”

That’s where things go wrong. A low p-value only tells you:

- The effect is unlikely due to random chance
- Given your assumptions
- Given your sample size
- Given your effect size

It does NOT guarantee:

- Better predictions
- Causal impact
- Business relevance
- Generalization to new data

Real-world trap: You add a new feature to a regression model:

🔹p-value = 0.001 

🔹Training error ↓

🔹Validation error ↑ 

What happened?

- The model learned noise, not signal
- Large datasets make tiny and useless   effects look “significant”
- Multicollinearity inflates confidence
- Leakage sneaks in quietly

A strong data scientist thinks like this:

Is this statistically significant and practically meaningful and stable out-of-sample?

Can we do other tests besides relying on this ?

If not, that beautiful p-value might be lying to you.

#datascience #pvalue

Security Data Scientist @ Microsoft | Data Science, Cloud Security, Machine Learning | Mentor

Intern Data Scientist:
“We can solve this using a deep neural network.”

Junior Data Scientist:
“Let’s build an AI system for this.”

Mid-level Data Scientist:
“Maybe a simple ML model with some features?”

Senior Data Scientist:
“Can we just use SQL for now?”

Because:

- Shipping a simple solution today beats a perfect model next quarter

Maturity in data science isn’t about using more AI. It’s about knowing when not to.

#datascience #machinelearning

Grateful, and truly thankful 🙏

Being recognized in Top 25 in 25 for AI/ML mentorship is special. This is not me, but it is representing everyone who trusted me, booked sessions, asked thoughtful questions, challenged ideas, and showed up ready to learn. Your support is the real reason behind this milestone.

I’m also honored to stand alongside so many talented mentors- some I deeply admire and personally follow. For me, this isn’t about which number I stand at, but where I stand- in a community driven by learning, generosity, and growth. 

Thank you for being part of this journey 🚀

Congrats to some of these wonderful people I follow:
Ajay Shenoy
Gabriela de Queiroz
Aishwarya Srinivasan
Nitin Aggarwal

#happyfriday #randomtopics

Breaking the recursion of "No Experience & No Job" in Data Science.

Ever feel stuck in the loop: I don't have experience to get a job. I don't have a job to build experience. 

 It's one of the biggest hurdles in data science. But there are few things you can do to break this recursion. Save this post, and please share with your network.

Volunteer for real projects: Contribute to open-source or community projects. Some of the popular platforms are:

- Datakind [https://lnkd.in/gmN4TpGv]
- Catch a fire [https://lnkd.in/dqYktzVd]
- Statistics Without Borders [https://lnkd.in/gpKmdj-x]
- Solve For Good [https://solveforgood.org/]
- Citizen Science [https://lnkd.in/g_6YMg-m]
- She loves Data [https://shelovesdata.com/]
- Northeast BigData Hub [https://lnkd.in/dfQNxT2T]

Build a strong portfolio: Showcase your skills with personal projects. Think end-to-end solutions: data cleaning, modeling, and deployment. GitHub is great. But lately I am finding this data-science portfolio site very simple, elegant and easy to set up: [https://lnkd.in/gyzy442S]

Network Network Network: Attend meetups, webinars, and local or virtual conferences. Build meaningful connections on LinkedIn, and don't shy away from reaching out to people in the domain/companies you inspire to be.

#datascience #volunteer #projects #networking #mentoring

undefined undefined

Mahammad Ameenulla

The 2026 Data Science Resume Checklist

🔹 The One-pager rule: Unless you have 10+ years of experience across multiple job transitions, keep it to one page. In fact it shows you have the communication skills to summarize complex work for stakeholders.

🔹 KISS (Keep It Stupid Simple): Avoid fancy templates, multiple columns, or infographics. The ATS prefer clean layouts to parse your skills correctly.

🔹 Lead with a Career Summary: Everyone cares about the TL;DR. Use 2-3 sentences to highlight your unique value proposition. 

🔹 Technical Skills at the Top: In 2026, skills-first hiring is the dominant standard. Organize it cleanly: Programming (Python, PySpark), Frameworks (LangChain, XGBoost), and Cloud (Azure ML, AWS SageMaker).

🔹 Impact Driven Bullets: What you did + How you did it + Business Impact : Your first three bullets must have measurable metrics (e.g., "Reduced triage time by 45%"). But don't force numbers everywhere; qualitative impact also matters, as long as it shows results- not only actions.

🔹 For New Grads: Build a strong and diversified project portfolio: If you lack work history, your projects are your experience. Showcase a range of tools (like GenAI/LLMs vs. Traditional ML) and domains (like Security vs. e-commerce) to prove your versatility.

I have created a sample resume, almost like a template to you help you. Find it attached with this post. Save it, and please share (repost) with your network.

#datascience #resume

Data Science Sample Resume

Technical Screening: Machine Learning Engineer Interview 🤖💻

I just completed a mock interview with my mentee for a Machine Learning Engineer role. I broke down session into two parts: Coding and general ML.

🔷 The Coding Challenge: Merging Cluster Telemetry

The Problem: You are given an array of intervals where intervals[i] = [start, end]. These represent "Busy" windows for a GPU cluster. Merge all overlapping intervals and return an array of the non-overlapping intervals that cover all the input intervals.

Sample Input: [[1,3], [2,6], [8,10], [15,18]]
Sample Output: [[1,6], [8,10], [15,18]]

The Nuance: While O (n log n) via sorting is the batch solution, we discussed using Interval Trees to handle real-time streaming data in O (log n) time.

🔷 5 Essential ML Breadth & Depth Questions

1. L1 vs. L2 Regularization: We broke down why L1 (Lasso) is your best friend for feature selection (it creates sparsity by zeroing out weights), whereas L2 (Ridge) is better for handling multicollinearity by shrinking weights evenly.

2. Generative vs. Discriminative Models: A fundamental shift. We compared modeling the boundary (P(y|x)) vs. modeling the actual data distribution (P(x,y)). Crucial for anyone working with modern LLMs. Follow up: Why computation is high in Generative?

3. The Bias-Variance Trade-off: How to move from a "low training error / high validation error" state. We discussed why Bagging (Random Forests) reduces variance, while Boosting (LightGBM/XGBoost) is the primary tool for reducing bias.

4. Metrics for Imbalance: Why Accuracy is a "vanity metric" in fields like fraud detection. We pivoted to Precision-Recall AUC to ensure the model actually provides business value. Follow up: Break it down into stakeholders' terms.

5. Responsible AI (The "Human" Element): How do you mathematically detect bias in a model? We looked at Demographic Parity and Disparate Impact-proving that "fairness" in 2026 is a technical requirement, not just a policy one.

Success in an MLE interview isn't about memorizing definitions, it's about explaining why one choice is better than another under specific system constraints.

Overall, he did great, I wish him all the best. 

#machinelearning #interview

Why SVM deserves more love 👀

• An unsung hero in ML
• Works great as one-class classifier for anomaly detection
• Handles non-linear data beautifully using the kernel trick
• Strong margin maximization → robust generalization (allows hard and soft margins- check out soft margin in SVM)
• Performs well even with small datasets
• Less prone to overfitting when tuned well

Not everything needs deep learning. Remember: Model explainability is inverse to its simplicity. 

While we all chase the GenAI and GPTs, 
sometimes, SVM still gets the job done quietly and effectively.

#datascience #machinelearning #SVM

LinkedIn

li_at

JSESSIONID

linkedin.com/feed/

Someone landing a data science job: "I got the sexiest job of 21st century."

Project 1: Let's start with some data cleaning and EDA, and the data:

- NULL
- NaN
- ""
- "unknown"
- "N/A"
- "0" (but sometimes meaningful!)

😂 

Takeaway: The cool job does come with some uncool stuffs, be mentally prepared. Modeling is just one piece, to get that right, you have to decide on lot of nuances and edge cases. BTW missing values are just one example.

#datascience #expectations #reality

I get a lot of inbox messages about paid content collaborations and sponsored posts.

For now, I’ve chosen to say no.

Not because they’re bad opportunities, but because of two personal reasons:

🔹I don’t want my content to turn into a sales pitch telling you which tool to use.

🔹I strongly believe in sharing unbiased, organic learnings with the data community.

Does this feel like a missed opportunity?
Honestly- yes, it does.

But I never started my LinkedIn journey to get paid. I started it to learn, share, mentor, and grow together as a community, and focus remains the same. 

To be clear- there’s absolutely nothing wrong with sponsored content. Many creators do it thoughtfully and add real value.

Everyone has their own perspective, and goals. This is just where I am right now- and I’m okay with that. 

#randomtopics #contents

graphical user interface, text, application

My notes on K-Means and  PCA

Once a while, over the weekend (as time permits), I used to mentor data science certification course from MiT in collaboration with Great Learning. 

There are many notes that I've created to share with learners for additional resources while covering the ML topics. This is the one I came across for K-Means and PCA. Thought of sharing it (just as raw as I had it). Hopefully all links are still relevant, if not please let me know. 

As I find more treasures, I will drop them 😊 

k-Means clustering: I find this article very helpful and well written.
https://lnkd.in/g3YX2c3v

K means vs. K-medoids clustering: Following article is very useful that I read for class. It also talks about K-means ++.
https://lnkd.in/g3EnJqn8

This wiki talks about the difference in K medians vs K medoids
https://lnkd.in/g9C3aXRX 

PCA: I found this article very useful. 
https://lnkd.in/gDtKz5ck 

Another Great Article on PCA:
https://lnkd.in/gU5TpCpz 

PCA: Little bit more deep dive with examples:
https://lnkd.in/g8mu2qE4 

PCA: Another good source: https://lnkd.in/gZXCKE7z 

Youtube video on PCA: If I have to pick one, it will be the one from Luis Serrano with example:
https://lnkd.in/g3e2i8vu 

Eigen Vectors and Eigen Values: This is one of the best videos I came across:
https://lnkd.in/g3YM2s9T 

Free book on Clustering available on research net site. Good to have it, to refer the topics as needed.
https://lnkd.in/gqV9jn82 

#datascience #machinelearning #kmeans

When he needs an ML refresher, then you know the 2026 interview market is getting serious 😊 

Let’s be honest: In a world of autonomous AI agents and "vibe coding," it’s easy to feel like you can fly blind. But the moment you step into a technical interview at a top-tier firm, the blindfold comes off.

You can’t prompt-engineer your way through a deep dive into gradient descent optimization, covariance matrices, or LLM quantization logic.

Whether you’re a world leader or a world-class engineer, fundamentals are the only things that don't change. If you’re preparing for interviews by just skimming headlines, you’re essentially doing- hoping for the best while missing the core details.

Why this matters in 2026:

🔹 Automation is high, but the bar is higher: Companies aren't hiring for syntax anymore; they are hiring for architectural intuition.

🔹 The Why > The How: Anyone can call an API. Very few can explain why the model is hallucinating or how to fix the data bottleneck.

🔹 No more black box prep: You need a roadmap that covers the 100 questions that actually get asked in the room.

Exactly why myself and Karun spent time to come up with our Interview Prep Book 📘 

We are currently running new year special- To help you kick off 2026 with total clarity, we’re offering our biggest discount yet:

✅ 26% OFF with the Code: NEWYEAR26 

🔗 Grab your copy here: [https://lnkd.in/gWGNMyt5]

Either end of Jan, or limited copies, its now or never :)

P.S. I have to give credit to Joe Reis. I came across one of his posts, and could not wait to put Gemini at work. Sorry Joe :) 

#datascience #interviews

📢 Job Opening in my sister team (Referral Possible)

Sharing a role from a team that I work closely with. If you meet the requirements and we’ve worked together professionally or through our mentorship, I’m happy to help with a referral.

Before reaching out, please read carefully:

- US Citizenship required (due to specific data access needs)
- Not a typical ML or Data Science role
- Not a junior role 
- Role is closer to service engineering with analytics exposure
- Data engineering skills are highly valuable
- Azure native tools experience is a big plus (Kusto (ADX), Logic Apps, Power BI, Synapse)
- Security experience will strongly differentiate you among hundreds of applicants

If all of the above checks out, feel free to connect and/or DM me.
If you are already MSFT employee and exploring change- I can help too. 

If not, please share this within your network- it might be exactly what someone else is looking for.

Let’s help the right people find the right opportunity🙌

Role: [https://lnkd.in/gSngDTMR]
#job #microsoft

I love to share messages like this, not to brag but to motivate others in the process. 

A mentee just signed an offer for a Senior Data Scientist role. 🎉

This is a second win for me in this year :)

Sharing this for you if you’re still in job hunt:

- Rejections don’t define your capability.
- Silence doesn’t mean failure. 
- Progress is invisible until it suddenly isn’t.

If you’re showing up, learning, refining your story, and staying consistent- your turn is coming.

And if you need a sounding board, resume feedback, interview prep, or just clarity on next steps- I’m always happy to help.

You don’t have to do this alone 💙

Keep going. The right “yes” is worth the wait.

#datascience #interview #offer

graphical user interface, text, application, chat or text message

Early-career me:
“I need to know everything.”

Today:
“I just need to know what matters.”

Talking to my mentors, I learned:
Experience doesn’t make me us louder.
It makes us simpler.

Stop chasing every new tool.
Start focusing on things that actually move the needle in right direction. 

Growth isn’t adding more.
It’s learning what to ignore.

#randomtopics #datascience

Not a data science post today …

please excuse my excitement 😆 

Patriots made it to SB after so many years 🏈 

Now cheering for Seahawks 🗣️

#randomtopics #nfl

Putting a Ferrari engine in a broken car won’t win you a race.

Same with data science.

If your data is messy, biased, incomplete, or wrong, no fancy models will save you, does not matter how large they are or how complex they are. 

This is one of the mistakes companies are making- slapping AI when data has nothing to offer. 

Algorithms are just methods. They don’t create signal. They only find patterns and amplify what’s already in the data- good or bad.

Real leverage comes from:

- Asking the right question
- Collecting the right data
- Cleaning it well
- Understanding its limitations

Model choice is very important.

I am not denying that. But data quality decides whether your model is useful or not.

Sometimes data is there, model is good. But there is no business significance. This is like ferrari engine in a nice car- but you are buying it for daily commute in a heavy traffic road. 

#datascience #machinelearning

Back at UC San Diego today, but this time I’m at the front of the classroom. 

I’m honored to join the Master of Data Science (MSDS) students as a guest speaker to discuss something I’m incredibly passionate about: Machine Learning in Detection Engineering.

As a veteran, I look at cybersecurity through a specific lens. In the field, an ambush happens when detection fails. In a network, a data breach happens for the exact same reason.

I’ll be walking the students through:

- The "Pyramid of Pain" (and why rules aren't enough anymore).
- Why a 99% accurate model can also be a nightmare in Security operations.
- The shift from static logic to AI Agents and behavioral modeling.

It’s a privilege to give back to the Triton community that helped sharpen my own toolkit. Let’s see if I can convince a few more Data Scientists to join the good fight in cybersecurity 🛡️💻

#datascience #security

Some days, it’s the quiet inbox messages that hit the hardest.

Knowing that someone genuinely looks forward to reading your posts, discusses them over coffee, and actually learns from them- that’s the best response I can hope to get for 

To everyone who follows, reads, reflects, and engages sincerely: this is what keeps me going.

Grateful beyond words. 🙏

I’ll keep showing up. For you ☕📚

#randomtopics #learning #mentoring

Logistic Regression: The Loss Function Everyone Uses but Few Truly Understand

Most people say: Logistic Regression uses log loss. But why?

Logistic Regression models probability, not classes. So instead of minimizing distance (like MSE in regression), it minimizes how wrong your predicted probability is. Thats the concept of entropy. 

That’s where Log Loss (Binary Cross-Entropy) comes in. I will let you check the exact equation, but will focus on intuition, we

• Predict 0.99 for a true class 1 → small penalty

• Predict 0.01 for a true class 1 → huge penalty

Log loss heavily penalizes confidently wrong predictions. That’s exactly what you want in classification.

Follow up: What Changes in Multi-Class Logistic Regression?

Now logistic regression predicts multiple class probabilities.

Instead of sigmoid, we use softmax. In other words, instead of binary cross-entropy, we use categorical cross-entropy.

What stays the same:
• Still probabilistic
• Still penalizes confident mistakes
• Still optimized via maximum likelihood

What changes:
• Loss sums over all classes
• Only the true class contributes to the loss
• Wrong class probabilities increase penalty indirectly

Interviews are geared to understand how strongly you understand fundamental concepts. 

#datascience #machinelearning

Stakeholder: we need to implement a Gen AI solution immediately.

Me: Great … What’s the use case?

Stakeholder: AI

Me: Right... but what specific problem are we trying to solve for our customers?

Stakeholder: The problem of not having AI.

Me: Got it. Then silently subscribing for a meditation app 😡🧘‍♂️ 😆 

The hype cycle is hitting a fever pitch. We’re reaching a point where "just add AI" is the new norm.

The hard truth? 

A state-of-the-art LLM won’t fix a broken data pipeline or a missing strategy. 

#datascience #humor #happyfriday

taplio

The Battle

Zach Wilson

Nirmal Budhathoki

Followers

Posts

Engagements

Likes

Performance ScoreZach Wilson's LinkedIn Performance

Performance ScoreNirmal Budhathoki's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Zach Wilson

Nirmal Budhathoki

Followers

Posts

Engagements

Likes

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Zach Wilson's LinkedIn Performance

Performance ScoreNirmal Budhathoki's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreZach Wilson's LinkedIn Performance