Deepika

Hey! I'm Deepika 👋 Recent CSE graduate learning AI/ML in public. Currently exploring NLP, sharing my learnings on Hashnode, and building real-world projects. Let's connect and learn together! #NLP #MachineLearning

Small win today: understood LangGraph persistence It’s basically state save + restore Which explains how chats magically resume Also did EDA on a dataset… Every month has all 4 seasons Real world data = creative writing? Or is this normal?

LangGraph is making agents feel less like code, more like systems. Built a small chatbot today Then went dataset hunting for XGBoost Funny part? Finding the right dataset > building the model Lowkey want a “describe → get datasets” tool

Got a Prompt Engineer interview tomorrow. Rounds include English test, aptitude, coding, GD… but nothing that evaluates prompting. Offer: 3 LPA + 2-year bond. Still attending though. Interviews are just another dataset for learning.

Today I learned XGBoost. Gradient Boosting but with: ⚡ speed 📉 regularization 🧠 smart optimizations Ran a small regression example. Tomorrow: real dataset. Starting to understand why everyone in ML interviews keeps asking about this. Respect to Tianqi Chen

Tried running Aden Hive today. Mistake #1: didn’t read the docs. Mistake #2: assumed it would “just work”. Eventually got it running on bash… but it quietly switched to an Anthropic model instead of the one I configured. AI agents are autonomous… apparently configs are too.

Learned Gradient Boosting for classification today. Same story as regression… just a different loss function. Also discovered ML involves more formulas than my memory likes. Took a break and tried some random photography too.

Today I studied the math behind Gradient Boosting. First 10 minutes: “This is easy.” Then came additive modeling… Then gradients… Now I’m confused but also curious. I guess solving problems is the only way forward. How did you guys actually learn this?

Spent today learning Random Forest. Thought it was a completely different algorithm… turns out it’s just many decision trees trained on random data. My Sunday ended with a realization (I’m dramatic). How’s yours?

Took an AI hiring challenge. Job title: AI Engineer Questions: • ML • NLP • Coding LLMs, agents, RAG… nowhere to be found. At this point I’m convinced AI interviews are just ML interviews in disguise.

For people running AI agencies: How do you figure out which companies might actually need your solution? Manual research? Signals you watch? Just experience? Trying to learn how agencies identify potential clients.

Building in AI feels strange sometimes. There’s always something new: new models new frameworks new terms Today I didn’t build anything. Just thinking, learning, processing. Tomorrow we build again.

Today: built a LangGraph agent with self-evaluation. Also today: discovered 5 new LLMs, 3 renamed ones, and 2 “state-of-the-art” replacements. Building agents is easier than tracking model names.

Worked on parallel workflows in LangGraph + built an IELTS Task 2 evaluator. Still haven’t locked in my March goals. Moving forward, figuring it out as I go.

Built an Employee Attrition Prediction system. Model + SHAP interpretability + FastAPI + Gradio + Docker. Production-ready setup. GitHub: github.com/Deepika081/emp… #LearningInPublic #MachineLearning

February so far: 2 projects done. Sentiment analysis, and a full employee attrition pipeline(EDA -> Deployment). Learned a lot, had fun, and honestly… it felt more like solving a mystery than coding ML.

Spent the day connecting FastAPI backend to a Gradio frontend for my attrition model. Prediction + probability + SHAP explanations working end-to-end. Feels good to see it live.

ML + SHAP = magic✨ My attrition model now tells not just who leaves, but why. App’s backend & frontend are next. What did you work on today?

@matthewncollins True. And just like TCP/IP, most people use it without understanding how powerful it actually is.

Prompts are today’s tcp/ip packets

Big realization today: Sklearn can build the model. But only you can understand the data. Spent more time on EDA → got better results. ML isn’t model-first. It’s data-first.

@0xAl3x_ Cool, I’m excited to see your mentor chatbot too.