Paul Iusztin's Linkedin Analytics

I’ve been working with GenAI for 3+ years. Here’s something all engineers must come to terms with: If you’re building LLM-powered applications, at some point, you’ll need to generate high-quality datasets to fine-tune SLMs. Why? → Fine-tuning SLMs reduces costs, latency, and throughput while maintaining high accuracy for specific tasks. → Some domains require specialized fine-tuning for better domain adaptation. → Fine-tuned models give you more control over AI behavior and response generation. That’s exactly what we’re tackling with our 𝗦𝗲𝗰𝗼𝗻𝗱 𝗕𝗿𝗮𝗶𝗻 𝗔𝗜 𝗔𝘀𝘀𝗶𝘀𝘁𝗮𝗻𝘁. ... and today, I’m breaking down the dataset generation feature pipeline we built for fine-tuning our summarization SLM. The input to our generation pipeline will be raw documents from MongoDB (Notion & crawled resources). And the output is a high-quality summarization dataset published to Hugging Face’s dataset registry. Since this pipeline generates features used to train an LLM, it’s called a feature pipeline. Here’s how it works, step by step: 𝟭. 𝗗𝗮𝘁𝗮 𝗘𝘅𝘁𝗿𝗮𝗰𝘁𝗶𝗼𝗻 → Pulls raw documents from MongoDB and standardizes formatting. 𝟮. 𝗗𝗼𝗰𝘂𝗺𝗲𝗻𝘁 𝗘𝘅𝗽𝗹𝗼𝗿𝗮𝘁𝗶𝗼𝗻 → Analyzes length & quality scores distributions to make informed decisions. 𝟯. 𝗗𝗮𝘁𝗮 𝗙𝗶𝗹𝘁𝗲𝗿𝗶𝗻𝗴 → Removes low-value content, keeping only high-quality documents. 𝟰. 𝗦𝘂𝗺𝗺𝗮𝗿𝗶𝘇𝗮𝘁𝗶𝗼𝗻 → We use a more powerful LLM (e.g., `gpt-4o`) to generate multiple summaries per document by varying temperature and sampling parameters (a process known as distillation) 𝟱. 𝗤𝘂𝗮𝗹𝗶𝘁𝘆 𝗖𝗼𝗻𝘁𝗿𝗼𝗹 → Filters out poor-quality summaries. 𝟲. 𝗗𝗮𝘁𝗮𝘀𝗲𝘁 𝗦𝗽𝗹𝗶𝘁𝘁𝗶𝗻𝗴 → Divides data into training, evaluation, and test sets (done before storing the dataset and not at training time!) 𝟳. 𝗩𝗲𝗿𝘀𝗶𝗼𝗻𝗶𝗻𝗴 & 𝗗𝗲𝗽𝗹𝗼𝘆𝗺𝗲𝗻𝘁 → Publishes the final dataset to Hugging Face. To keep the pipeline reproducible, trackable, and scalable, we manage it using ZenML, which: → Orchestrates the entire workflow from extraction to deployment. → Ensures traceability & versioning of pipeline runs & datasets. → Allows dynamic configuration for different filtering, summarization & structuring techniques. Even if you’re not deep into fine-tuning, at some point, you’ll need a structured way to generate datasets for specialized AI applications. This is one of the most critical components of your pipeline. Want to learn more? Check out the link in the comments.

The unthinkable happened... (And it wouldn't have been possible without you) The LLM Engineers Handbook GitHub repo has just crossed 2,800 stars. But that's not all... It's now less than 100 stars away from being the most popular repository in the Packt GitHub organization. Honestly, I'm flattered. This milestone wouldn't have been possible without your support. I’m truly grateful for every one of you who has contributed, starred, or even shared the repo. We’ve been working hard to make this repository a powerful resource for anyone looking to master LLMs. Along with the book, you'll learn how to implement production-ready, industry-level LLM & RAG applications. (Of course, following best practices in software engineering, system design and MLOps) And we've made some changes: - We’ve accepted PRs from contributors to fix bugs - Docs have been updated based on reader feedback If you haven't already, check it out. Contribute to the repo and help us hit that No. 1 spot. LFG🔥 (Link in the comments)

Over the past 6 months, I’ve seen the word “agent” thrown around a lot. But let’s be honest, most of the time it's misleading... In reality, what people often refer to as an "agent" is just Python scripts wrapped around LLM calls. (I’ve fallen into that trap myself early on.) But agents (and agentic systems) are fundamentally different from classic LLM workflows. Here’s how I break it down: 𝗪𝗼𝗿𝗸𝗳𝗹𝗼𝘄𝘀 → Follow a fixed, predefined sequence → Every step is planned in advance → Think “assembly line” logic—reliable, repeatable, predictable → Great for structured, stable tasks 𝗔𝗴𝗲𝗻𝘁𝘀 → Decide what to do next on the fly → Use reasoning to pave their path dynamically → Think of them like an intelligent assistant, dynamically adapting its plan → Ideal for unstructured, evolving tasks 𝗪𝗼𝗿𝗸𝗳𝗹𝗼𝘄𝘀 𝗮𝗿𝗲 𝗱𝗲𝘁𝗲𝗿𝗺𝗶𝗻𝗶𝘀𝘁𝗶𝗰. 𝗔𝗴𝗲𝗻𝘁𝘀 𝗮𝗿𝗲 𝗮𝗱𝗮𝗽𝘁𝗶𝘃𝗲. That’s the trade-off: Workflows = stability Agents = flexibility And that’s why true agents shine in complex GenAI apps where step-by-step logic won’t cut it. So next time you hear someone say “agent,” ask: Does it just call an LLM? Or does it reason, adapt, and act? Let’s raise the bar for what we call agentic systems.

AI/ML engineers, listen up! If you want to be in the top 1%, mastering ML frameworks isn’t enough... You need elite software engineering skills. Why? Because AI/ML systems don’t live in Jupyter notebooks. In production, they need: → Efficient, scalable code → Optimized cloud architectures → High-performance systems that can handle real-world demands One of my clients can attest to this first-hand... We rewrote parts of their **RAG ingestion pipeline in Rust** and reduced latency from **minutes to seconds.** That kind of performance boost doesn't come from just knowing how to prompt a model. … it comes from understanding how systems behave **under load**, how to optimize **I/O and memory**, and how to **build real software.** And that’s where 𝗖𝗼𝗱𝗲𝗖𝗿𝗮𝗳𝘁𝗲𝗿𝘀 comes in... CodeCrafters is one of the best platforms for leveling up real-world engineering skills. (I'm even planning to use it to learn rust) Here's why I say that: - It’s not just theory ↳ Want to understand how Git, Redis, Docker, Kafka, or Shell actually work? ↳You’ll build them yourself, piece by piece. - It forces you to write production-grade code. ↳ No hand-holding. No shortcuts. ↳ Just hands-on, real-world problem-solving that makes you a better engineer. - It’s perfect for AI/ML engineers who want to stand out. ↳ If you’re building LLM apps, RAG systems, or agentic workflows, you need strong SWE fundamentals. ↳ This platform helps you build that foundation. Interested? You'll get a 𝟰𝟬% 𝗱𝗶𝘀𝗰𝗼𝘂𝗻𝘁 if you use my affiliate link. (The link is in the comments)

I promised I'd make some bold moves this year. So, here goes nothing... On April 7th, I'll be giving my 𝗙𝗜𝗥𝗦𝗧 𝗶𝗻-𝗽𝗲𝗿𝘀𝗼𝗻 𝘁𝗮𝗹𝗸 at the QCon Software Development Conferences in London. Talk about biting the bullet! As an introvert (with a sprinkle of social anxiety thrown in), I must admit... I'm actually quite excited. ... and a little nervous - in the best way possible. The opportunity to share insights with an incredible community of engineers and data scientists doesn't come around often. Thus, I'll be sure to give it my all. In my talk, I’ll be diving into the data dimension of designing LLM and RAG applications. Namely, we will cover: - The data flow and features pipeline that powers LLM and RAG systems - LLMOps best practices for ensuring data is shared, versioned, processed, and analyzed for LLM training and inference. - Concrete use cases like building an LLM Twin and a Second Brain AI Assistant. The goal: To give you a clear, practical framework for architecting the data layer of LLM systems that scales with accuracy, reliability, and efficiency. If you'd like to attend, I'll leave a registration link in the comments (with £60 off the price code) See you there 👊

If you're an aspiring AI engineer, listen up... There are 2 basic skills you must master before touching any ML model: 𝟭. 𝗣𝗿𝗼𝗴𝗿𝗮𝗺𝗺𝗶𝗻𝗴 (Python is a good start) 𝟮. 𝗖𝗹𝗼𝘂𝗱 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 Without these skills, you will not even be able to deploy or scale your AI/ML models effectively in production. Fortunately, 🎧 Eric Riddoch noticed this problem and filled the need - Enter 𝗖𝗹𝗼𝘂𝗱 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗳𝗼𝗿 𝗣𝘆𝘁𝗵𝗼𝗻 𝗗𝗲𝘃𝗲𝗹𝗼𝗽𝗲𝗿𝘀 live course. I’ve had the pleasure of knowing Eric for almost a year, and I can confidently say he’s a brilliant cloud, DevOps, and MLOps engineer. He’s also an incredible teacher with a knack for making complex topics approachable and fun to learn. Here’s what you’ll learn in the course: - Enterprise-level AWS account management - Fundamentals of cloud engineering - Designing cloud-native RESTful APIs that scale from 4 to 4 million requests/day - Writing, testing, locally mocking, and deploying code using AWS SDK and OpenAI - Advanced observability and monitoring techniques: Logs, metrics, traces, and alerts If you're concerned that the course is not with your desired cloud vendor, don't worry... The principles learned during the course can easily be transferred to other platforms (e.g., GCP and Azure) I strongly recommend this course to anyone looking to lay the foundation for a career in AI, ML, and MLOps engineering. The next cohort starts on March 31st and runs until May 16th! Using the code DECODINGML will get you 10% off your registration. And if that's not enough, Eric offers a scholarship program that can significantly reduce the price - depending on your use case. Check out the link in the comments to enroll.

If you think you “know” prompt engineering... Think again. I’ve been following Nir Diamant for a while now - his GitHub repos and Substack have become go-to resources for AI practitioners. He has a rare gift: The ability to break down complex GenAI topics like he’s teaching a 7-year-old (without dumbing anything down). ... And now he’s done it again with a new eBook: 𝗣𝗿𝗼𝗺𝗽𝘁 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗳𝗿𝗼𝗺 𝗭𝗲𝗿𝗼 𝘁𝗼 𝗛𝗲𝗿𝗼 – 𝗠𝗮𝘀𝘁𝗲𝗿 𝘁𝗵𝗲 𝗔𝗿𝘁 𝗼𝗳 𝗔𝗜 𝗜𝗻𝘁𝗲𝗿𝗮𝗰𝘁𝗶𝗼𝗻 This isn’t just another “use more bullet points in your prompt” kind of guide. It’s a practical deep dive with: → Code examples → Real-world exercises → Clear explanations of common mistakes → And the subtle mechanics behind great AI interaction One reader put it best: “𝘞𝘰𝘸... 𝘐 𝘵𝘩𝘰𝘶𝘨𝘩𝘵 𝘐 𝘬𝘯𝘦𝘸 𝘱𝘳𝘰𝘮𝘱𝘵 𝘦𝘯𝘨𝘪𝘯𝘦𝘦𝘳𝘪𝘯𝘨 𝘣𝘦𝘧𝘰𝘳𝘦 𝘳𝘦𝘢𝘥𝘪𝘯𝘨 𝘵𝘩𝘪𝘴 𝘣𝘰𝘰𝘬, 𝘣𝘶𝘵 𝘐 𝘸𝘢𝘴 𝘭𝘪𝘨𝘩𝘵𝘺𝘦𝘢𝘳𝘴 𝘢𝘸𝘢𝘺 𝘧𝘳𝘰𝘮 𝘦𝘷𝘦𝘯 𝘣𝘦𝘪𝘯𝘨 𝘢𝘸𝘢𝘳𝘦 𝘰𝘧 𝘵𝘩𝘦 𝘯𝘶𝘢𝘯𝘤𝘦 𝘢𝘯𝘥 𝘤𝘰𝘮𝘱𝘭𝘦𝘹𝘪𝘵𝘪𝘦𝘴... 𝘛𝘩𝘪𝘴 𝘪𝘴 𝘵𝘩𝘦 𝘣𝘰𝘰𝘬 𝘵𝘰 𝘱𝘪𝘤𝘬 𝘶𝘱 𝘪𝘧 𝘺𝘰𝘶 𝘸𝘢𝘯𝘵 𝘵𝘰 𝘨𝘦𝘵 𝘱𝘢𝘴𝘵 𝘫𝘶𝘴𝘵 𝘬𝘯𝘰𝘸𝘪𝘯𝘨 𝘸𝘩𝘢𝘵 𝘱𝘳𝘰𝘮𝘱𝘵 𝘦𝘯𝘨𝘪𝘯𝘦𝘦𝘳𝘪𝘯𝘨 𝘪𝘴, 𝘢𝘯𝘥 𝘴𝘵𝘢𝘳𝘵 𝘮𝘢𝘴𝘵𝘦𝘳𝘪𝘯𝘨 𝘵𝘩𝘦 𝘤𝘳𝘢𝘧𝘵.” I couldn’t agree more. If you’re building LLM apps, agentic systems, or even just playing with ChatGPT prompts, check this eBook out. You’ll walk away with a new toolkit that goes far beyond theory. Want a copy? The link is in the comments.

Here's something most people overlook: The RAG feature pipeline is the most important part of the entire AI stack. Not the LLM. Not the prompt. Not even the fancy agent framework. Why? Because if your retrieval sucks, your generation will too - no matter how good your model is. Let me walk you through how we architected our RAG feature pipeline for the 𝗦𝗲𝗰𝗼𝗻𝗱 𝗕𝗿𝗮𝗶𝗻 𝗔𝗜 𝗔𝘀𝘀𝗶𝘀𝘁𝗮𝗻𝘁: The pipeline ingests raw documents from MongoDB, where both Notion and crawled content are stored in a single standardized collection. We don’t care about the source - just that the data is clean and usable. The output? → Chunked + embedded documents → Stored in a dedicated vector store → Indexed and ready for semantic search Where does this pipeline fit? It’s an offline batch pipeline, decoupled from the live user experience. At query time, the pipeline does not run. All processing is done beforehand, so retrieval is fast, stable, and cost-efficient. Meanwhile, the online pipeline (our agentic RAG module) fetches the chunks, reasons over them, and generates the answer. The architecture is made up of 7 key components: 𝟭/ 𝗗𝗮𝘁𝗮 𝗘𝘅𝘁𝗿𝗮𝗰𝘁𝗶𝗼𝗻 Pulls all raw documents from MongoDB, regardless of source. 𝟮/ 𝗗𝗼𝗰𝘂𝗺𝗲𝗻𝘁 𝗙𝗶𝗹𝘁𝗲𝗿𝗶𝗻𝗴 Applies quality scores to drop noisy or low-value docs. 𝟯/ 𝗖𝗵𝘂𝗻𝗸𝗶𝗻𝗴 Splits documents into manageable segments for vectorization. 𝟰/ 𝗣𝗼𝘀𝘁-𝗣𝗿𝗼𝗰𝗲𝘀𝘀𝗶𝗻𝗴 Applies one of two retrieval strategies (chosen from our YAML configs): Parent Retrieval → Links each chunk to its full doc Contextual Retrieval → Adds summaries to enrich semantic relevance 𝟱/ 𝗘𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴 All chunks (for both strategies) are vectorized using a configurable embedding model (OpenAI or Hugging Face) 𝟲/ 𝗜𝗻𝗱𝗲𝘅𝗶𝗻𝗴 Embeddings are stored and indexed in MongoDB for fast lookups. 𝟳/ 𝗙𝗶𝗻𝗮𝗹 𝗢𝘂𝘁𝗽𝘂𝘁 A structured, searchable knowledge base—ready for RAG-based generation. The entire pipeline is managed by ZenML, which is: → Reproducible → Configurable → Versioned → Traceable If you’re serious about building production-grade GenAI systems, this is where you focus. ... because 90% of your generation's quality is determined before the LLM even gets involved. Let’s stop treating pipelines like afterthoughts. They are the product. Want to learn more? Check the link in the comments.

RAG has kept “dying” for the past 4 years. But here’s why that will never happen: All the LLMs (even the most advanced ones) struggle without the right context. It doesn’t matter if your model has 128k+ token windows or cutting-edge fine-tuning... If it doesn’t retrieve the right data or the context is full of noise or formatted incorrectly, it won’t generate the right answers. That’s why retrieval is the hardest part of RAG. Most RAG failures aren’t about generation - they happen before the LLM even sees the data. If the retrieval step is weak, your AI assistant will: - Fetch irrelevant information - Miss critical details - Hallucinate confidently wrong responses But more context isn’t the answer... Better context is. Lesson 5 of the Second Brain AI Assistant course is all about fixing retrieval with a production-ready RAG feature pipeline that. (And it's now live!) Namely, in this lesson, you will learn: - The fundamentals of RAG. - How to design and implement a production-ready RAG pipeline - Implement contextual retrieval (an advanced RAG technique) from scratch. - Implement parent retrieval (another advanced RAG technique) using LangChain. - Extend LangChain to add custom behavior using OOP. - The critical role of chunk size in optimizing retrieval quality - Write a configuration layer to switch between different algorithms and models dynamically - How to manage everything with an MLOps framework (we use ZenML) By the end of this lesson, you’ll be equipped with the skills to build a flexible, modular RAG feature pipeline. This pipeline gives access to our AI assistant to our Second Brain and provides reliable context to generate meaningful answers. Sounds interesting? Pick up lesson 5 today. (The link is in the comments) Thank you, Anca Ioana Muscalagiu, for contributing with another fantastic lesson to @Decoding ML !

Over the past 2 years, I've been on a mission... (And it's finally coming together) My goal has been to give as much back to the AI community as possible. I've done this by creating a total of 6 open-source courses designed to help others build real-world, production-grade AI systems. Now, I'm thrilled to announce that I've brought all my FREE courses together in one place. These are not your typical tutorials showing isolated concepts in scattered Notebooks... They're designed for those looking to master production and industry-level AI. It covers everything from: - End-to-end GenAI, LLM, RAG apps - MLOps/LLMOps for AI scalability - AI Systems for real-world applications I've been sure to constantly refine and improve these courses to keep them at the cutting edge. Our latest project, the 𝗦𝗲𝗰𝗼𝗻𝗱 𝗕𝗿𝗮𝗶𝗻 𝗔𝗜 𝗔𝘀𝘀𝗶𝘀𝘁𝗮𝗻𝘁 course, is our crown jewel so far - It's an end-to-end project that takes you through the full process of building a production-level AI assistant. The world of AI is moving incredibly fast. ... I believe that knowledge-sharing is key to pushing the field forward. I’m currently working on a new open-source course focused on AI agents that I know you’ll love. (More on that soon) All in all, I'm incredibly proud of what we've built so far. I'm excited to see how these resources help others in their AI journey. Check it out, share it, and let's keep building the future together. The link is in the comments.

There's a major problem with using general-purpose LLMs to build apps. But not enough people are talking about it... Of course, I'm talking about: → Skyrocketing API bills → Vendor lock-in → Degrading performance we can't do anything about → Lack of control of your data Thankfully, I've found a solution. In Lesson 4 of the 𝗦𝗲𝗰𝗼𝗻𝗱 𝗕𝗿𝗮𝗶𝗻 𝗔𝗜 𝗔𝘀𝘀𝗶𝘀𝘁𝗮𝗻𝘁 course, we tackle these challenges head-on. Specifically, you'll learn: - How to fine-tune open-source small language models (SLMs) using LoRA and QLoRA to specialize them on your tasks. - Efficient fine-tuning with tools like TRL, Unsloth AI, and Comet - Architecting modular, scalable pipelines with MLOps in mind for production-ready applications. - Deploying your own specialized models on Hugging Face Inference Endpoints as a real-time API endpoint. - Evaluating your models using vLLM to ensure they’re production-ready. By the end of this lesson, you'll have the tools and knowledge to take back control of your AI systems (without breaking the bank or being tied to a single provider) Ready to get started? The link is in the comments.

AI engineers must understand this to build any successful system: (Especially when working with RAG, LLMs, or agent-based apps) How to architect data pipelines. We all know data is the lifeblood of the systems we build. You can have the most advanced algorithms, but without clean data, you're just spinning your wheels. To understand how data pipelines for LLM apps should look, let’s architect one that powers our Second Brain AI Assistant. Here are the core components: 𝟭/ 𝗗𝗮𝘁𝗮 𝗰𝗼𝗹𝗹𝗲𝗰𝘁𝗶𝗼𝗻 Data sources can be many things - Notion, crawled links, custom APIs... you name it. In this case, we'll use Notion’s API to collect personal data, extract all the links, and standardize everything into Markdown format. Why Markdown? Because the internet is full of it. Thus, LLMs are heavily trained using this format and know how to work with it. To replicate steps down the line, we make a snapshot of the raw data into a data lake like S3. 𝟮/ 𝗘𝗧𝗟 𝗱𝗮𝘁𝗮 𝗽𝗶𝗽𝗲𝗹𝗶𝗻𝗲 This is where the magic happens. After collecting the data, we perform several transformations: 1. Download the Notion docs from S3 2. Load them in memory into Pydantic objects (content + metadata) 3. Crawl all the links inside the Notion docs (when building the AI assistant, we want to generate answers based on the content of the links as well → this is where the power of our lists of saved resources kicks in) 4. Compute a quality score per document using a combination of heuristics and LLMs (as LLM calls are expensive, you first want to squeeze the most out of it using heuristics) 5. Store all the documents and the quality score inside a document database (with this, we can compute statistics and filter documents down the line based on our needs) For our setup, we’ve used MongoDB: a Swiss knife to store our unstructured documents and, later on, the embeddings for RAG. We can think about it as part of our logical feature store. 𝟯/ 𝗠𝗟𝗢𝗽𝘀 𝗙𝗿𝗮𝗺𝗲𝘄𝗼𝗿𝗸 To keep everything organized and maintainable, we use ZenML This helps us easily manage, version, and deploy our offline data pipelines. 𝗧𝗮𝗸𝗲𝗮𝘄𝗮𝘆: Architecting and managing these pipelines is crucial for creating LLM/RAG systems. If you're an AI/ML engineer stepping into the GenAI world, get comfortable building and managing data pipelines that crawl unstructured data and process it for LLMs. (Yes! Along with the 1000+ things you must know, AI Engineers must also dig into data engineering.) Want to see how to do this? Check the link in the comments.

Less than 3 years ago, I graduated with my master’s in ML. Fast forward to today, I’m: - Contracting - Consulting - Creating content Globally! So, what helped me accelerate my career? 𝟭. 𝗜 𝘀𝘁𝗮𝗿𝘁𝗲𝗱 𝗲𝗮𝗿𝗹𝘆 While still in university, I didn’t wait for the “perfect” opportunity. By my third year, I had already landed a junior SWE role and was working like crazy. 𝟮. 𝗜 𝗽𝘂𝘁 𝗶𝗻 𝘁𝗵𝗲 𝘁𝗶𝗺𝗲 - 𝗻𝗼 𝘀𝗵𝗼𝗿𝘁𝗰𝘂𝘁𝘀. For years, I spent 10-12 hours a day researching, building, and creating- mostly alone. Monday to Saturday. That helped me transform 3 years of experience into 6. 𝟯. 𝗜 𝗮𝗰𝗰𝗲𝗽𝘁𝗲𝗱 𝘁𝗵𝗲 𝘁𝗿𝗮𝗱𝗲-𝗼𝗳𝗳𝘀 I knew that accelerating my career meant sacrifices in other areas. I had less time for social life. ... but this meant more time for deep work and learning There was only one exception: My girlfriend. No matter how much I worked, I made time for her - she kept me sane. 𝟰. 𝗛𝗲𝗮𝗹𝘁𝗵 > 𝗛𝘂𝘀𝘁𝗹𝗲 → Workouts: 3-4x per week → Food: Eat clean, avoid junk → Sleep: 7-8 hours, no compromise → Stress: Keep it under control (no toxic jobs, no endless pressure) 𝟱. 𝗡𝗼 𝗮𝗹𝗰𝗼𝗵𝗼𝗹. 𝗡𝗼 𝘀𝗺𝗼𝗸𝗶𝗻𝗴. Clarity is everything. You can’t afford to be foggy-minded when making big moves. (I still enjoy exceptions here and there, but nothing that slows me down.) 𝟲. 𝗜 𝗳𝗼𝗹𝗹𝗼𝘄𝗲𝗱 𝗺𝘆 𝗶𝗻𝘁𝘂𝗶𝘁𝗶𝗼𝗻 - 𝗮𝗹𝘄𝗮𝘆𝘀. I ignored the “safe” paths. Regardless of what “more experienced” people told me, I always made hard decisions. Most people project their fears onto you, but they don’t live your life. 𝟳. 𝗜 𝘁𝗼𝗼𝗸 𝘁𝗵𝗲 𝗺𝗶𝗻𝗱𝘀𝗲𝘁 𝗼𝗳 𝗮𝗻 𝗲𝘅𝗽𝗹𝗼𝗿𝗲𝗿 If you follow paved paths, you’re in high competition. If you carve your own path, you move faster. I never aimed to be top 1% in existing tracks. I created my own track instead. 𝟴. 𝗜 𝗻𝗲𝘃𝗲𝗿 𝘀𝗲𝘁𝘁𝗹𝗲𝗱 When my managers said: - “The job market is tough.” - “This is what others get paid.” - “You’re asking for too much.” - “Have patience.” I didn’t listen. I doubled down and kept pushing forward. It’s exhausting at times, but curiosity > comfort. And that’s what kept me going. So, if you’re early in your career, here’s my advice: → Be intentional about how you spend your time. → Build skills relentlessly - don’t wait for permission. → Challenge conventional wisdom - not all “advice” is useful. → Prioritize health - burnout kills long-term success. → Bet on yourself - no one else will do it for you. Happy Sunday!

Many ML projects fail to transition from POC to production-ready. Here's one simple reason why: The initial focus was never on scalability or production constraints. Fraud detection presents a perfect use case for building production-first ML systems. It combines the need for real-time and batch processing, low-latency predictions, and high-accuracy models. Here's a look into what that system could look like: 𝗗𝗮𝘁𝗮 𝘀𝗼𝘂𝗿𝗰𝗲𝘀 In fraud detection, you deal with real-time transactions, streaming data and historical records. Real-time and streaming data require you to instantly compute features, while historical records help track user profiles and spending patterns. In our architecture, real-time features are computed through HTTP requests, streaming data will flow through Kafka topics, with historical data stored in a data warehouse for batch processing. 𝗙𝗲𝗮𝘁𝘂𝗿𝗲 𝗽𝗹𝗮𝘁𝗳𝗼𝗿𝗺 At the heart of the system is the feature platform, like Tecton ... This centralizes all the features. More specifically, it allows us to manage features in an offline store (for training - high throughput) and online store (for serving - low latency). Using the same feature engineering logic during training and inference avoids training-serving skew. 𝗙𝗲𝗮𝘁𝘂𝗿𝗲 𝗽𝗶𝗽𝗲𝗹𝗶𝗻𝗲𝘀 These convert raw data into meaningful features. By centralizing your features into a feature store like Tecton, you can leverage their feature views to define features once and reuse them across models and pipelines. A feature view is defined as a data source(s) + a function that maps raw data into features. Next, using a process known as materialization, you sync the raw data sources with the online/offline stores while applying the transformations. 𝗧𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗽𝗶𝗽𝗲𝗹𝗶𝗻𝗲𝘀 It ingests features and labels to train models (stored in a model registry). Leveraging the feature store, you can easily apply time-traveling strategies to version your dataset. 𝗜𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝗽𝗶𝗽𝗲𝗹𝗶𝗻𝗲𝘀 It takes new transaction data, enriches it with features from Tecton's feature platform, and applies the trained model to generate predictions. Online stores are crucial at serving time. They provide low-latency access to up-to-date features. When a transaction occurs... The pipeline quickly retrieves pre-computed features, combines them with real-time features, and computes predictions. 𝗢𝗯𝘀𝗲𝗿𝘃𝗮𝗯𝗶𝗹𝗶𝘁𝘆 𝗽𝗶𝗽𝗲𝗹𝗶𝗻𝗲 Lastly, an observability pipeline is essential for monitoring the system's health and detecting drifts. The final touch is an alarm system that sends emails or SMS or denies transactions if fraud is detected. Want to dive deeper into building such systems? Check out the link in the comments.

The course is finally complete! You can learn for free how to build your Second Brain AI assistant with LLMs, agents, RAG, fine-tuning, LLMOps, and AI systems techniques. Thanks for sharing Andrew Davidson is the crown jewel of Decoding ML

RAG is evolving. Standard retrieval is no longer enough... If you’re building LLM-powered applications, you need something more powerful: 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗥𝗔𝗚. That’s exactly what we’re tackling in Lesson 6 of the Second Brain AI Assistant course. ... and it's now LIVE! Most RAG systems passively retrieve context and hope the LLM generates the right response. But this approach is limited. → What if your assistant needs to reason about multiple sources? → What if retrieval alone doesn’t fully align with the user’s intent? → What if the retrieved context isn't enough, and the system needs to iterate? Agentic RAG bridges that gap. Instead of just fetching documents, agents dynamically plan, retrieve, and refine their responses, making AI assistants: → Smarter → More interactive → More accurate And in Lesson 6, we’re building one from scratch... Specifically, you’ll learn: → How to build an Agentic RAG module that goes beyond simple retrieval → Integrating retrieval with AI agents for dynamic, multi-step reasoning → Adding LLMOps, such as prompt monitoring to optimize retrieval workflows for cost, latency, and throughput → Evaluating long and complex LLM chains to ensure reliability (moderation, hallucination, response performance) → Scaling retrieval architectures to handle real-world AI assistant demands By the end of this lesson, you’ll understand what it takes to build stable, efficient, and intelligent RAG-powered assistants. Huge thanks to Anca Ioana Muscalagiu for contributing another fantastic lesson to Decoding ML! The link for Lesson 6 is in the comments :)

If you're building GenAI apps, you need to understand one thing: LLMs don’t “think” like humans. They generate responses based on probabilities. ... and how you craft your prompts can make or break your application. That’s why I’ve been diving deep into "𝗣𝗿𝗼𝗺𝗽𝘁 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗳𝗼𝗿 𝗟𝗟𝗠𝘀." I must say, it’s one of the best resources out there for practitioners building with LLMs. → It’s written with character → It focuses on how to actually build GenAI products (practical theory) → It’s packed with fundamentals that every AI/ML engineer should know I especially loved how they teach prompt engineering beyond basic techniques you can find everywhere. They teach you how to manage and process the context injected into the prompts, which is required to scale LLM apps beyond PoCs. I’m still working my way through the book, but after finishing the first 7 chapters, I can confidently say it's a must-read. Specifically for: → Anyone building GenAI applications → Engineers looking to optimize their prompt engineering skills → Experienced AI practitioners looking to build modular LLM apps. I'd also like to say a huge thanks to John Berryman for sending me a copy - I'm absolutely loving it. If you’re serious about building with LLMs, this book is worth your time. (I'll leave a link to it in the comments)

I’ve just started working at a new AI start-up (still in shadow mode) on building cutting-edge products using GenAI, LLMs, RAG, and Agents. All the good stuff. This will help me improve my content with even more real-world battle-tested content and stories. So excited to start providing more in-depth insights from this move 🤟

Cursor + Claude 3.7 keeps disappointing me. I am not a vibe coder. I strongly articulate what should be implemented, how it should be implemented, and the issues within the codebase. But when things get tricky and require extensive understanding of the code, infrastructure and data, it just doesn’t work. I lost 2 hours of my life talking to Claude (Extended) when a GitHub Issues thread solved it in 5 minutes. My take is that LLMs are great for quickly ramping up your project (instead of digging into the documentation), but you still have to: - design - plan - know the best engineering practices - be open to use Google What is your experience?

I can’t get enough of this meme 😂. Are you down that path yet?