Lessons learned as an AI/ML Engineer in 2023

When I first heard “Generative AI,” I thought it was only about ChatGPT. But it’s much more than that 🤖 It’s about teaching machines to create. Text, images, code — even ideas. So I started learning step by step 📘 First — Python basics 🐍 Then — Machine Learning foundations 📊 Then — how LLMs actually “think” 🧠 Every small concept felt like a superpower 💡 Prompt engineering blew my mind. It’s not magic — it’s math, logic, and creativity combined. Now, I’m exploring real-world applications on GCP ☁️ Building AI tools that actually help people. If you’re curious about AI, start today. Small steps can lead to something big 🚀 Follow me to see how I turn GenAI into real-world projects! 🤝 #GenerativeAI #ArtificialIntelligence #MachineLearning #AIProjects #DataScience #FutureOfWork #CloudAI #LearningJourney #TechInnovation

Recently, I started learning something called Model Context Protocol (MCP) — and honestly, it’s one of the most exciting things I’ve come across in AI so far! In simple words, MCP is a new way to let AI systems talk to custom tools, apps, and data. Think of it like this 👇 ➡️ Just like how a phone connects to different apps, MCP lets AI models connect to your own tools. You can create your own AI-powered services — for example, an app that manages expenses, answers database queries, or controls smart devices — and then make it usable directly inside AI platforms like ChatGPT. I’m using a library called FastMCP, which makes it super easy to build and test these AI servers in Python. — just the beginning! Excited to keep exploring and eventually build something practical. If you’ve never heard of MCP — it’s definitely worth checking out. It’s going to change how AI systems integrate with real-world tools. #AI #FastMCP #MCP #Python #OpenAI

🚀 RAG vs CAG - The Next Evolution in Generative AI! 🤖⚡ We’ve all heard about RAG (Retrieval-Augmented Generation) - but have you met its powerful next-gen cousin, CAG (Context-Augmented Generation) ? 👀 Let’s break it down 👇 💜 RAG (Retrieval-Augmented Generation) RAG is like a smart librarian 🧑🏫 It retrieves external documents 📚, filters the noise 🧹, and uses them to generate accurate answers 💬 💡 Example: You ask ChatGPT: What are the latest features in PySpark 3.5? → RAG fetches data from documentation, blogs, or APIs, then summarizes it to answer you accurately ✅ 💜 CAG (Context-Augmented Generation) CAG is like a context-aware research assistant 🧠 It doesn’t just retrieve info + it understands your intent, merges domain knowledge, and keeps context consistent across sessions 🔄 💡 Example: You’re building a data pipeline with PySpark, and the model already knows your data schema, prior errors, and config. → CAG remembers this context, merges it with documentation, and gives you personalized, context-synced code suggestions ⚙️ No re-explaining. No repetitive prompts. Just continuous, context-rich intelligence 🪄 🎯 In short: 🧭 RAG = Retrieve → Reason → Respond 🧬 CAG = Understand → Synchronize → Generate The future of AI isn’t just about finding answers - it’s about staying contextually aware 🌐 💥 RAG is retrieval. CAG is real understanding. 👇 Check out the visual flow below to see the difference! Artificial Intelligence News Python Google NVIDIA OpenAI #AI #GenerativeAI #RAG #CAG #LLM #LangChain #LangGraph #ArtificialIntelligence #MachineLearning #DataEngineering #DataScience #PySpark #Innovation #AITrends #ContextAugmentedGeneration #RetrievalAugmentedGeneration 🚀🤖💡

Learning Data Science ≠ Learning AI In 2025, everyone wants to “learn AI 🤖” But very few actually understand what they’re trying to learn. As a Data Science student at UC Irvine, I’ve realized something: - You don’t need to “learn AI.” - You need to learn how to think with data. - You can memorize every Python library, every ML model, every prompt-engineering trick… but if you can’t ask the right question, the model will only give you noise. So lately, I’ve been focusing less on tools and more on mindset: 🧠 framing problems clearly 📊 understanding bias in data 🗣️ communicating insights like a storyteller, not a statistician Because in the age of ChatGPT, Gemini, and Claude...tools will evolve every month. But the ability to think, reason, and question? That never gets outdated. Maybe the real skill isn’t data science at all, It’s scientific thinking in a world drowning in data. If you are more curious, definately check "Thinking with Data ~ Max Shron" Curious to know, do you think universities are teaching this mindset well enough? #DataScience #ArtificialIntelligence #MachineLearning #StudentJourney #UCI #InternationalStudent #TechCareers #AIethics

From Code to Context My Ongoing Journey with Large Language Models (LLMs) Working with LLMs has reshaped the way I think about backend engineering and data systems. What began as Python-based microservice development soon evolved into building retrieval-augmented generation (RAG) pipelines and context-aware AI applications that go far beyond simple API calls. At NJIT, my research focused on applying transformer architectures and semantic retrieval to extract insights from unstructured data. Later, integrating LangChain, FAISS, and AWS Bedrock helped me bridge classical data pipelines with intelligent reasoning, enabling models to connect facts, understand context, and produce domain-specific outputs with precision. The most exciting part? The shift from “what” a system does to “why” it does it. LLMs introduce a new engineering challenge orchestrating memory, context, and trust. It’s no longer just about clean APIs and scalable infrastructure; it’s about making AI systems explainable, adaptive, and auditable. I see a future where AI engineers and data engineers converge, working together to design pipelines that allow models not just to respond but to reason. LLMs aren’t replacing software logic, they’re enriching it with understanding. #LLM #GenerativeAI #LangChain #AWSBedrock #RAG #AIEngineering #Python #MachineLearning #NLP #SoftwareDevelopment #DataEngineering #AI

Artificial Intelligence isn’t replacing developers - it’s redefining them. The next generation of engineers won’t just code; they’ll design intelligent systems that learn, reason, and adapt to challenges. AI development demands creativity, logic, and data-driven thinking - making developers the architects of the intelligent future. The transformation begins by learning the skills that power modern AI. To pivot into AI, here’s the roadmap: - Master the foundations of Python - Learn the core concepts of ML - Build small but complete projects - Explore Deep Learning and LLMs - Deploy and integrate real-time AI apps At Reliable Software, we believe the future developer is an AI engineer at heart. #AI #MachineLearning #DeveloperJourney #ReliableSoftware

What excites you in AI Agents? - question to fellow ML Engineers AI Agents feel more like Backend Engineering to me rather than ML - Develop agents using frameworks (LangGraph, PydanticAI, LLamaAgents), connect internal/external tools, create infra, deploy the solution, actively monitor. Obviously a very simplistic picture, but the "predictive" and "data" part is mostly missing. Going deeply with agents feels to me like following the "wrong" route if you want to grow as a Machine Learning Engineer.

🚀 Today’s AI/ML pulse: faster models talking to models & the new 2025 coding-LLM scoreboard. 1️⃣ Cache-to-Cache (C2C) lets LLMs share KV-cache states directly—no text, no tokens, no waiting. Early tests show 30-50 % latency drop & 2× throughput for mult ti-agent or edge-cloud chains. One-line code swap, outsized efficiency gains. Paper walk-through: https://lnkd.in/dm_QdxwG  2️⃣ The 2025 “who codes best” list is out. Benchmarks now weigh not just pass@k but debug speed, repo-level reasoning & cost per PR. Spoiler: open-weight models s are closing the gap on frontier ones. Full comparison: https://lnkd.in/dNGV42P8 Bottom line: smarter inter-model chatter + sharper coding specialists = cheaper, faster AI software teams. Time to rethink your stack. 💡 #AI #MachineLearning #LLM #GenerativeAI #SoftwareEngineering #EdgeAI #MultiAgentSystems

Understanding LLMs - The core of Gen AI. Before we dive into building with Gen Al, it's important to understand what powers it. To truly grasp how Generative Al works, we first need to understand LLMs - Large Language Models. So what are they, really? At a high level, an LLM is a deep learning model trained on massive amounts of text data (books, code, articles, etc.) to understand and generate human-like language. It doesn’t “know” facts the way humans do — it predicts the next most likely word based on patterns it has learned from data. Let’s break it down 👇 🧩 Architecture: Most modern LLMs are built on the Transformer architecture (the same one introduced by Google’s 2017 paper “Attention is All You Need”). ⚙️ Training: Models learn by processing billions of text tokens and adjusting internal weights to capture relationships between words and context. 📦 Inference: When we query a model (e.g., via an API), it doesn’t retrieve data — it generates a new response by predicting token by token. 🔍 Applications: Code generation, text summarization, chatbots, content creation, or even backend automation with natural language prompts. ⚠️key Takeaways LLMs are not “magic boxes.” They’re just powerful probabilistic models that understand and generate text by recognizing patterns in language. In upcoming posts, I will dive deeper into the core concepts of AI #java #genai #javadeveloper

😩 Tired of just using LLMs? ⚡️ It's time to build them. 😉 Forget the black box. This incredible GitHub repo is your golden ticket to coding a ChatGPT-like LLM in PyTorch – from scratch! If you've ever wanted to truly understand Generative AI, this is your definitive, step-by-step guide to implementing a ChatGPT-like LLM in PyTorch from scratch. This isn't a high-level overview; it's the definitive, hands-on roadmap to mastering the core mechanics of Generative AI: The Roadmap Highlights: 1. Foundation: Implement the core GPT architecture, including the attention mechanism, from the ground up. 2. Zero Dependencies: It uses pure PyTorch—no external LLM libraries—ensuring you understand every single line of code. 3. Full Cycle: Covers the entire LLM workflow, from working with text data to pretraining on unlabeled data and instruction fine-tuning (the "alignment" step). 4.Accessibility: The code is designed to run on a conventional laptop, utilizing a GPU if available. This is the official code companion for the book "Build a Large Language Model (From Scratch)." If you are a Data Scientist, ML Engineer, or anyone serious about understanding the future of AI, this repository is required reading. Stop relying on API calls and start controlling your own destiny. 🔗 The Ultimate Resource: https://lnkd.in/dSUXyR-g #AIAgents #GenerativeAI #LLMs #DeepLearning #PyTorch #DataScience #MachineLearning #AIdevelopment #Coding