"Learn Experiment Tracking with MLflow from Azumo"

𝑾𝒉𝒆𝒏 𝑶𝒍𝒅-𝑺𝒄𝒉𝒐𝒐𝒍 𝑴𝑳 𝑴𝒆𝒆𝒕𝒔 𝑴𝒐𝒅𝒆𝒓𝒏 𝑮𝒆𝒏𝑨𝑰 🤖 A few days back, during our tea break with Aniket & Anurag, my two teammates, we were having a good discussion on GenAI and its impact across industries.. At one point, the discussion turned to whether 𝐭𝐫𝐚𝐝𝐢𝐭𝐢𝐨𝐧𝐚𝐥 𝐌𝐋 𝐚𝐥𝐠𝐨𝐫𝐢𝐭𝐡𝐦𝐬 are still relevant since everything can seemingly be done better with 𝐥𝐚𝐫𝐠𝐞 𝐥𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐦𝐨𝐝𝐞𝐥𝐬, but Anurag made an interesting point: even with the recent advancements in GenAI, traditional ML algorithms will always remain relevant because they've laid the groundwork for modern AI applications. This really caught my attention, and I started thinking about a project where modern GenAI techniques and traditional ML could be combined to solve a problem efficiently. This weekend, I got the chance to experiment with just that! Played with unstructured image data - used Snowflake 𝐌𝐮𝐥𝐭𝐢𝐦𝐨𝐝𝐚𝐥 𝐄𝐦𝐛𝐞𝐝𝐝𝐢𝐧𝐠𝐬 + 𝐏𝐂𝐀 + 𝐊-𝐌𝐞𝐚𝐧𝐬 to turn a folder of images into meaningful clusters… and yes, no labels were harmed in the process!🐱🐶 Check out the full blog here 👉 https://lnkd.in/dNA8HUG2 #dataengineering #snowflake #embedding #unstructured #clustering

Attending the Snowflake World Tour in Chicago this week made one thing clear — the boundaries between data engineering, AI, and analytics are disappearing fast. Snowflake’s new Cortex AI ecosystem shows what happens when reasoning, context, and data governance live in the same place. Imagine asking a question in natural language and Snowflake automatically knows which tables, metrics, and logic to use — that’s where we’re headed. With Notebooks, Cortex Agents, and Semantic Views, the platform is becoming a full AI environment for the modern data team. The next era of analytics won’t just be “data-driven.” It’ll be context-aware and autonomous. #Snowflake #CortexAI #DataScience #AI #Analytics #Innovation #Datassential

This observation is often shared on LinkedIn. I would like to think that the vast majority of people entering this field have this awareness. If not then course providers and authors of books need to do a better job. Familiarity with the MLOps process should guide you through this. If you practice with real world data as the source then none of this will come as a surprise. One area where I do believe there is a problem is with universities. The question often posed when you propose a project is where do we source the data from. It should be part of the challenge to curate the raw data, understand the domain and go through the process of cleaning and refining the information before it is fit for training models. Transforming messy data into a clean data set is often seen as a mundane, repetitive step but in reality this is where critical thinking needs to be applied. Going through the process aids familiarity with the features, domain knowledge and semantics. All of this is essential when you come to train the models.

Struggling to Pick the Right AI Algorithm? This Cheat Sheet Breaks It Down by Use Case! From Text Analysis to Image Classification, Anomaly Detection to Recommender Systems — get clarity on what works best, where. Perfect for: AI Engineers, Data Scientists, ML Beginners Save this post — your next project will thank you. To get complete guide for free: 1. Connect with me 2. Like this post 3. Comment “AI” below, and I’ll send it to you! Pdf credit goes to respective owner. Follow Pratham Chandratre for more!

Just wrapped up a CNN project on the CIFAR-100 dataset — 100 challenging object categories in 60,000 images. The process wasn’t smooth at first 😅 Faced repeated FileNotFoundError issues when trying to access the dataset meta file. Solved it by dynamically detecting the correct dataset path and automating label loading. Implemented EarlyStopping for stable training and achieved 39.44% test accuracy. It’s not state-of-the-art yet, but a great baseline for experimenting with data augmentation and transfer learning. Key takeaway: understanding your data pipeline is just as important as optimizing your model. #DeepLearning #TensorFlow #CIFAR100 #MachineLearning #DataScience #AI here's a github link to the project: https://lnkd.in/dUi8_9Y7 and some images. Constructive criticism is welcome

AI Engineering Progress — Day 2 AI vs ML vs Data Science — What’s the Difference? Everyone talks about AI, Machine Learning, and Data Science... But what actually separates them? 👇 🧵

🚀 Curious how to make a pre-trained AI model specialize in your own data? In my latest video, I demonstrate fine-tuning in Amazon Bedrock with a full hands-on walkthrough.   📌 Topics covered: What fine-tuning is and why it matters Preparing a dataset in JSONL format and uploading your dataset to Amazon S3 Creating a fine-tuning job in Bedrock Setting up inference: on-demand vs provisioned throughput Testing the custom model in the playground Interpreting validation results from S3   💡 Video link is in the first comment — check it out for the full demo!   #AI #MachineLearning #AmazonBedrock #FineTuning #CustomModels #DataScience

🚀 Curious how to make a pre-trained AI model specialize in your own data? In my latest video, I demonstrate fine-tuning in Amazon Bedrock with a full hands-on walkthrough.   📌 Topics covered: What fine-tuning is and why it matters Preparing a dataset in JSONL format and uploading your dataset to Amazon S3 Creating a fine-tuning job in Bedrock Setting up inference: on-demand vs provisioned throughput Testing the custom model in the playground Interpreting validation results from S3   💡 Video link is in the first comment — check it out for the full demo!   #AI #MachineLearning #AmazonBedrock #FineTuning #CustomModels #DataScience

Over the past few months I’ve been experimenting with multimodal AI on Databricks. With Mosaic AI Model Serving now able to accept multimodal inputs and a growing lineup of vision‑capable foundation models like Claude Sonnet and Llama 4, we can finally process images alongside text through the same API and vector search infrastructure. That said, I still get asked one question all the time: Should we convert images into text and embed them, or embed the images directly? Here’s how I’ve been thinking about it. ✅ Use image → text → embedding when: - You care about cost efficiency and interpretability. A vision model like Claude 3.7 can describe colors, objects and context in plain language, and a text embedding model can vectorize those descriptions. - Your domain already has rich text (e‑commerce catalogs with standardized product photos, internal docs). - You’re iterating on a workshop, demo or proof‑of‑concept and want to keep the pipeline simple. 🎯 Use image → embedding when: - You need true multimodal retrieval where visuals carry meaning text can’t capture. Databricks can now host vision models directly, and there are great third‑party options like Cohere’s Multimodal Embed 4, Nomic‑Embed, Meta ImageBind and CLIP. - Your use case is visual search (fashion, design, medical imaging), or you expect users to upload images without any accompanying text. - You want cross‑modal search: typing a query and retrieving matching images from a vector index. 🔧 The Hybrid approach: In production I often combine the two. Use a vision model to generate structured descriptions, then embed those descriptions. It’s the best of both worlds: real image understanding, interpretable features and lower costs at scale. Building multimodal pipelines is no longer research—it’s part of my day‑to‑day work on Databricks, and it’s changing how we build search and recommendation systems. I would love to hear how others are approaching this. #Databricks #AI #VectorSearch #Multimodal #MLops #DataIntelligence

💡 Time Complexity of 10 Popular Machine Learning Algorithms Understanding the runtime behavior of ML algorithms isn’t just for theory geeks — it’s a practical skill every ML engineer and data scientist should master. Here’s why 👇 ✅ It helps you choose the right algorithm for your dataset size. ✅ It builds core intuition about model efficiency. ✅ It prevents wasted hours (or days) training infeasible models. For example: 🚫 Support Vector Machines (SVMs) or t-SNE become impractical on large datasets due to their polynomial time complexity. ⚠️ Ordinary Least Squares (OLS) regression grows cubically with the number of features, making it unsuitable for high-dimensional problems. ✅ Linear models with SGD or Naive Bayes scale beautifully for massive datasets. 🧠 The chart below neatly summarizes the training and inference complexities of 10 common ML algorithms — from Linear Regression to K-Means. 📊 Key takeaway: When you know how an algorithm scales, you can design smarter, faster, and more efficient ML pipelines. #MachineLearning #DataScience #DeepLearning #AI #MLOps #BigData #ComputationalComplexity #LearningAlgorithms #AIEngineering