Dipu Sardar - দিপু সরদার

21/08/2026

16/08/2026

👉Successfully Complete my first ML project called "Ford Car Price Prediction"
The steps:
-> EDA
-> Test Split
-> Encoding
-> Model & Evaluation

The final result is,
R2 = 83%
Adjusted R2 = 83% ✌️

14/08/2026

Day 2 of core ML. Today was linear regression — not the formula, but why the formula.

The chain that finally clicked:

→ A best fit line is just a guess: y = mx + c
→ The intercept (c) is where that line starts when x = 0
→ Residual error is the vertical gap between each real point and my line
→ Square all those gaps, average them — that's the cost function (MSE)
→ Plot cost against different values of m, and you get the cost curve — a valley
→ Gradient descent is how the model walks down that valley to the lowest point

The insight I didn't have yesterday: the model isn't drawing a line. It's minimising a number. The line is just a side effect.

Notes below — written by hand, because typing them doesn't make them stick.

01/08/2026

Just learned about Orphan Processes in OS 👀

When a parent process dies before its child, the child doesn't disappear — it becomes an orphan, still running, reassigned to init (PID 1) by the kernel.

The interesting part?

→ Orphan ≠ Zombie. Orphan is alive. Zombie is dead but stuck in the process table.
→ getppid() returning 1 means you're an orphan.
→ Daemons like nginx and sshd are intentional orphans — that's how background services actually work.

OS concepts hit different when you run the code and watch PPID change live.

27/07/2026

Currently working through Applied Statistics & Queuing Theory this semester, and I keep coming back to a simple truth: the math behind AI/ML isn't optional — it's the foundation.
Today's focus: Hypothesis Testing with the Z-Test.
A quick breakdown from my notes:
→ Formulate H₀ (null) and H₁ (alternative)
→ Set your significance level (α)
→ Calculate Z = (x̄ − μ₀) / (σ/√n)
→ Compare against critical values
→ Reject H₀ if |Z| > Critical Z
It's easy to jump straight to scikit-learn and skip this step, but every model's confidence intervals, p-values, and A/B testing logic traces back to concepts like this. Slowing down to actually understand why the math works has made the ML side click a lot faster for me.
Currently on Module 2 (Math for AI) of my 6-month AI Engineering roadmap. One concept at a time. 🚀

25/07/2026

Forked my first process today 🍴
Lab work on Operating Systems — implementing fork(), handling parent/child processes, and cleaning up zombie processes with wait(). Simple concept, but a solid building block for understanding how modern OSes juggle multiple processes.

15/07/2026

This is one of the most advanced LLM/Agent tool nowadays.
In my case, I've used a lot this platform for making my study and work easier.

-The free version of Claude is not enough, every time it is showing me the timeline, because I finish my daily token limit when I get it.

Some tips to control token,
-Try to use low effort for easy task. ↘️↘️
- Use multiple account for long work. ✌️

This two tips will help you a lot and thanks me later. 😇

15/07/2026

Compiler Design Sessional Lab — and I'm hooked already.

There's something genuinely beautiful about compiler design. It's the layer where human-readable code quietly transforms into something a machine can execute — lexical analysis breaking source code into tokens, parsing building syntax trees, semantic analysis catching meaning errors, all the way down to code generation. Every language you've ever used — Python, C, JavaScript — exists because someone built this pipeline.

It's the class where you stop just using programming languages and start understanding how they're built.

📚 Resources that are helping me get started:
- Compilers: Principles, Techniques, and Tools (the "Dragon Book") — the classic, dense but worth it
- Neso Academy's Compiler Design playlist on YouTube — great for building intuition before the theory
- GeeksforGeeks Compiler Design section — quick reference for concepts like FIRST/FOLLOW sets, LL(1) parsing
- Flex & Bison official docs — essential once you start building

⚙️ Setting up the toolchain (Flex + Bison):

On Mac:
brew install flex bison
Then verify with flex --version and bison --version. Xcode Command Line Tools (xcode-select --install) should already give you gcc/clang to compile the generated C code.

On Windows:
- Easiest path: install WSL (Windows Subsystem for Linux), then inside it run:
sudo apt update && sudo apt install flex bison gcc
- Alternative: install MinGW-w64 or Cygwin if you want to stay native without WSL.

First lab project incoming 👨‍💻 — starting with a simple lexical analyzer. Excited for where this course takes me.

06/07/2026

🧠 Why every AI/ML Engineer should be on Medium
Medium isn't just a blog platform — it's one of the most valuable tools in an AI engineer's toolkit.
Here's what it gives you:
📄 Research Digests — Complex papers (GPT-4, BERT, diffusion models) explained in plain language
💻 Real Tutorials — End-to-end Python, PyTorch, and scikit-learn projects you can actually use
🚀 Personal Brand — Writing builds credibility. Your Medium + GitHub = your real portfolio
📡 Industry Trends — Stay ahead on LLMs, RAG, agents, and what's actually shipping
🤝 Community — Connect with 700K+ AI/ML writers on publications like Towards Data Science
💰 Earn While Learning — The Medium Partner Program pays you for every read
My personal rule: Read 1 Medium article per day.
In 6 months, you'll know what others took years to figure out.
Are you on Medium? Drop your profile below 👇

03/07/2026

Before you train a model — you need to understand your data.
And that starts here. 👇

🔷 Measures of Central Tendency
These answer: "Where is the center of my data?"
➤ Mean (μ) — The average. Simple, but sensitive to outliers.
➤ Median (M) — The middle value. Robust, and preferred for skewed data.
➤ Mode (Mo) — The most frequent value. Essential for categorical features.

📐 Measures of Spread
These answer: "How spread out is my data?"
➤ Range = Max − Min → Total spread
➤ IQR = Q3 − Q1 → The middle 50%, and far more reliable than range alone.

🚨 Detecting Outliers with Fences
Once you know your IQR, you can spot extreme values:
➤ Upper Fence = Q3 + 1.5 × IQR
➤ Lower Fence = Q1 − 1.5 × IQR
Any value outside these fences? That's an outlier — flag it before modeling.

📦 The 5 Number Summary
Remove outliers first. Then compute:
.min()2Q1 (25th %ile)np.percentile(arr, 25)3Median / Q2np.percentile(arr, 50)4Q3 (75th %ile)np.percentile(arr, 75)5Maximumarr.max()
💡 This summary is the backbone of your box plot — one of the most powerful tools for EDA (Exploratory Data Analysis).

✅ The order matters:
1️⃣ Explore the distribution
2️⃣ Detect & handle outliers
3️⃣ Compute the 5 Number Summary
4️⃣ Visualize with a box plot
5️⃣ Then move to modeling

These are foundational concepts — but they're skipped far too often in rushed ML pipelines. Master this, and your feature engineering gets dramatically better.
🔁 Repost if this helped someone on your feed.
💬 Drop a comment: which stat trips you up the most?