STACKERCODE’s Post

Just published a new article: Mastering Binary Tree Operations: A Comprehensive Guide. Binary trees are one of the most foundational data structures in computer science and ML, powering everything from database indexes to expression evaluators. In this piece, I break down the core operations every developer should know, including traversal techniques (inorder, preorder, postorder), recursive and iterative search, insertion and deletion using the transplant method, and the elegant Euler Tour traversal that captures all three classical walks in a single pass. I also walk through a practical example: evaluating a mathematical expression tree step by step using the Euler Tour algorithm. Whether you are preparing for technical interviews, brushing up on fundamentals, or designing more advanced algorithms, this guide should give you a solid reference point. Read the full article here: https://lnkd.in/dqAeU9TQ #DataStructures #Algorithms #ComputerScience #SoftwareEngineering #Programming

I'm starting a new blog series on Medium, and I'd love for you to follow along. The series is called "Building with Data", a collection of honest, practical write-ups on LLM engineering, ETL automation, RAG pipelines, and what it actually takes to go from a Data Science degree to building real AI systems. No toy datasets. No theoretical overviews. Just the real stack, the real decisions, and the things that broke before they worked. Here's what's coming: - From Masters in Data Science student to LLM engineer: what my curriculum missed and what actually got me hired. - How I automated 80% of manual reporting with Python + Power Automate (the exact stack). - Building a RAG pipeline on SEC 10-K filings and what actually broke. - Agentic RAG vs classic RAG: which one should you actually build in 2026? - And more throughout the year. I'll be posting regularly, and every piece will be built around something I actually built or shipped (not surface-level advice). If you're a Data Science/Machine Learning student, an early-career engineer, or someone navigating the gap between academic ML and production AI systems, this series is for you. First post is live now. Link in the comments. See you later! #DataScience #MachineLearning #LLM #Python #DataEngineering #AIEngineering #CareerGrowth #RAG #ETL #WebScraping #RealData #Pipeline #CICD #CFBR #NLP #Hire #Masters #MS #AgenticAI #AgenticRAG

🔹 Part 4/6 – Feature Engineering ⚙️ This is where raw data became model-ready intelligence. After EDA, I built a feature engineering pipeline to capture time dependencies, trends, and environmental interactions. ⏰ Time Features: Hour, day, and month encoded using sin/cos transformations to preserve cyclic patterns. 🔁 Lag Features: AQI lags (1, 3, 6, 24 hours) extracted from MongoDB historical data to capture temporal memory — ensuring no data leakage. 📊 Rolling Features: Moving averages and standard deviations (6h, 24h) to model trends and volatility. ⚡ Change Feature: Past AQI change to detect rising or falling pollution trends. 🔗 Interactions: Temperature × humidity and wind decomposed into X/Y components for real-world relationships. 🧠 Key focus: time-series consistency + leakage prevention 📌 This step transformed raw AQI data into meaningful features, enabling better learning and more accurate predictions. #AQIPrediction #DataScience #MachineLearning #AI #Python #DataAnalytics #MLOps #10Pearls #TechPakistan #FeatureEngineering #TimeSeries #DataPipeline #MongoDB #Forecasting #DataEngineering #RealWorldProjects

Beyond Point-Predictions: Probabilistic Stock Forecasting 📈 Standard ML models often struggle with market volatility because they provide a single "best guess." I built a Bayesian Time-Series System that prioritizes uncertainty quantification—providing the probability distributions necessary for risk-aware financial decisions. 🔗 Live API:https://lnkd.in/gCEJpkxB 🔗 Portfolio: https://lnkd.in/gF7kbKwv The Engineering Stack Models: SARIMA, XGBoost, LSTM, and Bayesian Regression. The Bayesian Edge: Used Bayesian A/B testing and confidence intervals to quantify model "certainty." Production-Ready: Built with FastAPI, tracked with MLflow, and containerized via Docker. Frontend: Streamlit dashboard for real-time uncertainty visualization. Key Engineering Wins Shifted from monolithic notebooks to modular production code. Solved time-series challenges like non-stationarity and data leakage. Implemented a full MLOps lifecycle from data ingestion to cloud deployment. special thank to Krish Naik Mayank Aggarwal Monal S. I’m looking for feedback from the #QuantFinance and #MLOps communities—how are you handling uncertainty in your production pipelines? #MachineLearning #DataScience #Bayesian #Finance #MLOps #FastAPI #Python #AI

I analyzed 10 of the world’s largest public companies using Python and a set of core fundamental metrics commonly used in equity research. Here’s what the data revealed 👇 🟢 NVIDIA — Composite Score: 0.80/1.0 55.6% profit margin, 65.5% revenue growth, and Debt/Equity of 0.07. The strength here isn’t narrative-driven — it’s supported by exceptional profitability, growth, and capital structure. 🟡 Alphabet, Meta, Microsoft — clustered between 0.45–0.48 These companies demonstrate consistent performance across key dimensions: strong margins, stable growth, and disciplined leverage. 🔴 Tesla — did not meet screening criteria Revenue declined -2.93% YoY, with a P/E ratio of 348. The combination of slowing growth and elevated valuation led to early exclusion. Project overview: A quantitative screening pipeline built to evaluate companies across profitability, growth, liquidity, leverage, and valuation. Tech stack: • Python (Pandas) — data processing and feature engineering • DuckDB — querying large-scale Parquet datasets efficiently • Hugging Face — access to Yahoo Finance historical data (~100M+ rows) • Matplotlib — multi-panel analytical dashboard Pipeline: Data ingestion → Ratio computation → Rule based filtering → Composite scoring → Visualization Key takeaways: • No single metric is sufficient in isolation A company can appear strong on one dimension and weak on another holistic evaluation is critical. • Profitability + growth + low leverage consistently signal strength High performing companies tend to balance all three rather than optimize for just one. • Valuation must be contextualized with growth High P/E ratios without supporting growth trends introduce significant risk. • Data engineering is as important as modeling Handling large scale financial data efficiently was a core challenge in building this pipeline. This project reinforced a key principle: well structured data combined with thoughtful metric selection can provide clear, actionable insights into company fundamentals. 🔗 Full implementation and documentation available on https://lnkd.in/gcYFFVg7 #QuantFinance #DataScience #Python #StockMarket #Investing #FinTech #MachineLearning

unpopular opinion: stop obsessing over the algorithm. i see people spending weeks comparing models. XGBoost vs LightGBM. Random Forest vs SVM. tuning hyperparameters for hours trying to squeeze out that extra 1% accuracy. and their dataset has 500 rows. 😭 here's what nobody wants to hear — your model is only as good as your data. full stop. i learned this the hard way. i spent days building a "better" model for my ATM cash forecasting project. tried everything. and then i went back and looked at the features more carefully, added a few meaningful variables, cleaned some noise i had ignored. accuracy jumped more in 2 hours than it did in 3 days of model tuning. more data. cleaner data. better features. that's the boring answer. that's also the right answer. the algorithm is the last 10%. the data is the other 90%. but somehow courses spend 80% of the time on algorithms and 20% on data quality. i think a lot of ML beginners (including past me) fall into this trap because tuning models feels like progress. it feels technical. it feels impressive to say "i tried 7 different models." collecting better data and cleaning it properly? that feels like grunt work. it's not glamorous. but that's where the real results are. fight me in the comments 👇 #MachineLearning #DataScience #MLengineering #Python #OpenToWork

Recommender Systems using recbole #machinelearning #datascience #recommendersystems #recbole RecBole is a unified, comprehensive and efficient framework developed based on PyTorch. It aims to help the researchers to reproduce and develop recommendation models. In the lastest release, our library includes 94 recommendation algorithms [Model List], covering four major categories : General Recommendation Sequential Recommendation Context-aware Recommendation Knowledge-based Recommendation We design a unified and flexible data file format, and provide the support for 44 benchmark recommendation datasets [Collected Datasets]. A user can apply the provided script to process the original data copy, or simply download the processed datasets by our team. Features : General and extensible data structure We deign general and extensible data structures to unify the formatting and usage of various recommendation datasets. Comprehensive benchmark models and datasets We implement 94 commonly used recommendation algorithms, and provide the formatted copies of 44 recommendation datasets. Efficient GPU-accelerated execution We design many tailored strategies in the GPU environment to enhance the efficiency of our library. Extensive and standard evaluation protocols We support a series of commonly used evaluation protocols or settings for testing and comparing recommendation algorithms. https://lnkd.in/gpdpgmwm

There are two types of Data Scientists in the tech industry today: 🥉 Type 1 → knows how to import libraries They know how to import pandas and run model.fit(), but when the dataset gets too large, or the pipeline slows down, they pause and aren't sure what's happening under the hood. 🥇 Type 2 → knows the time complexity of the algorithm under the hood They understand memory limits, Big-O time complexity, and exactly why a nested loop will crash the server. Now, personally, I have never enjoyed DSA; in fact, every time I saw a LeetCode problem, it felt like useless academic math that had nothing to do with building real Machine Learning models. But over the past 4 weeks, I decided to face it and have been grinding NeetCode every single day, solving 3 problems daily, and it has fundamentally rewired my brain into mastering two crucial skills: [1] Problem Framing: When a stakeholder gives you a messy, chaotic business problem, DSA trains your brain to map it to a logical structure instantly. For example, routing patients through the hospital isn't a spreadsheet problem; it requires Graph Traversal. [2] Problem Decomposition: Hard DSA problems are impossible to solve all at once; you have to break them down into base cases, helper functions, and edge cases, which is exactly how you have to build scalable, production-grade Machine Learning pipelines. Code is becoming commoditised by AI, and today, anyone can ask LLMs to write a Python script. Still, AI cannot optimise an enterprise system if the human prompting doesn't make it understand the underlying architecture. Do you want to be a Type 1 or a Type 2 Data Scientist? #DataScience #SoftwareEngineering #MachineLearning #NeetCode

I built an open-source, multi-agent research assistant that runs entirely on a 1GB cloud VM. Managing a PhD application workflow takes a lot of time. Between reading papers, tracking deadlines, and organizing notes, I wanted an always-on AI partner. But most agent frameworks require heavy local servers or expensive APIs. I built Aurora to solve this. She is completely free to run and incredibly lightweight. Instead of forcing a single heavy model to do everything, Aurora uses multi-model routing via OpenRouter’s free tier. She routes complex tasks (like summarizing research methodologies) to models like Hermes 3 405B, and quick commands to fast models like Gemma 12B. Beyond typical chat, she handles my daily workflow directly from Telegram: • Deep Research: Queries the ArXiv API for papers and scrapes live webpages in parallel using asyncio. • Vision & Images: Analyzes screenshots, graphs, or photos instantly using Gemini Flash. • Deadline Tracking: Syncs with my local trackers, Notion, and Google Calendar to send proactive warnings. • Daily Life: Logs my expenses by category, tracks daily habits, and sends me a scheduled morning briefing at 8 AM. • Server Monitoring: Runs real bash commands to report on RAM, disk space, and active processes. To keep it running under 500MB of RAM, I bypassed heavy vector databases and LangChain. She uses persistent semantic memory built entirely on SQLite and pure-Python cosine similarity. In the demo video below, Aurora checks her own live Linux server stats via bash, confirming she is running on just 467MB of RAM. If you are a researcher looking for an automated assistant, or an engineer interested in building low-resource multi-agent systems, the entire project is open-source. Architecture and Code: https://lnkd.in/gvuzKgpm #ArtificialIntelligence #AI #MachineLearning #ML #Python #OpenSource #PhDResearch #AgenticAI #SoftwareEngineering

Spent the last few weeks replicating "Physics of Language Models: Part 3.1" by Allen-Zhu & Li (https://lnkd.in/ednVbBY8), a very interesting paper with no official code release. The core finding of this work: memorization is not the same as knowledge extraction. A model can achieve near-perfect loss during pretraining and still answer 0% of factual questions about the data it has seen. Getting from storage to extraction requires specific training conditions: data augmentation (diversity, permutation...) and/or mixed training (i.e. pretraining with QA-like data). 🖥️ Code is open: https://lnkd.in/e4eBeP6p If you've replicated other results from the Physics of LLMs series or want to contribute additional conditions, pull requests are welcome.