Vartika Sharma

50% of my college time was spent learning ML. I specifically focused on generative models and computer vision. Even after college, I worked for 2 years in computer vision. I had very little experience with NLP. But then it took me just 6 months to cover classical NLP, NMT,and attention, and start working with language models like BERT, GPT etc. And my learning philosophy remained the same. Learn in-depth maths, write code and explore libraries. This speed-up was almost like transfer learning for me. This is true not only for ML but also for other fields. Be it backend, frontend, or devops. As you learn more, you’re able to connect the dots and fill in information gaps faster. You don’t need to acquire all the information to understand a concept since things are more or less the same or follow a similar framework.

544

13 Comments

Day 1: Multi-LLM Workflows In the evolving landscape of AI, Multi-LLM workflows are becoming a cornerstone of scalable and efficient applications. These workflows combine multiple Language Model calls to optimize performance, balancing accuracy, latency, and cost. On Day 1 of this series, we explore three foundational workflows—Prompt Chaining, Parallelization, and Routing—that serve as building blocks for sophisticated AI systems. ✅ Key Insights from Day 1: 1. Prompt Chaining    This workflow decomposes complex tasks into sequential steps. Each step processes a subtask, feeding its output as input to the next. This enhances accuracy, making it ideal for tasks requiring precision, such as generating and validating technical documents. 2. Parallelization    By processing independent subtasks concurrently, parallelization significantly reduces latency. This workflow is best suited for large-scale tasks like batch processing customer reviews or analyzing stakeholder impacts. 3. Routing    Routing leverages input-specific strategies to dynamically select the most appropriate prompt or model. This ensures specialization and efficiency, especially in applications like customer support ticket triaging or handling diverse document types. ✅ Practical Applications:  - Extracting structured insights from raw data through prompt chaining.  - Accelerating data analysis via parallel processing.  - Automating task assignment using intelligent routing mechanisms. These workflows highlight the flexibility and power of Multi-LLM systems in tackling diverse challenges. Day 1’s tutorial not only outlined these concepts but also included practical steps for implementing a prompt-chaining workflow in a Google Colab notebook. From setting up dependencies and configuring APIs to executing workflows, the tutorial provides a comprehensive foundation for developers looking to explore multi-LLM architectures. ✅ Looking Ahead  These foundational workflows open the door to a world of possibilities. Whether optimizing for precision, speed, or specialization, Multi-LLM workflows can transform how AI applications are built and scaled. Stay tuned for the next steps in this journey as we dive deeper into advanced multi-LLM strategies. For detailed sample code, check out the Google Colab notebook(https://lnkd.in/g6QPgAFN) #ArtificialIntelligence #GenerativeAI #MachineLearning #NaturalLanguageProcessing #MultiLLM #PromptEngineering

54

Looking for a roadmap for NLP + LLMs? This one CMU course should do it for you Go from basics like word representation and language modeling to fine tuning, prompting, RAG, all the way to tool use, complex reasoning and multilingual NLP. You will also see RL and mixture of experts here. Link in comments

1,216

36 Comments

Hot off the press, "Foundations of Large Language Models", published on January 16th, 2025, is the most updated and comprehensive guide to the advancements in Generative AI (Gen AI) and Large Language Models (LLMs). Packed with insights, frameworks, and the latest research, this book is a must-read for staying ahead in the AI revolution. 🔑 Why This Book is Unmissable? 1️⃣ The Most Recent Research: Published in January 2025, it incorporates state-of-the-art frameworks, methodologies, and findings. Explores breakthroughs in alignment strategies, prompting methods, and scaling laws. 2️⃣ Deep Focus on the Gen AI Era: Details how LLMs have redefined AI workflows across industries. Explains how pre-training, fine-tuning, and human feedback alignment work together to build powerful AI systems. 3️⃣ Updated Frameworks and Techniques: Advanced Prompting: Chain-of-thought reasoning, automatic prompt optimization, and self-refinement methods. Alignment Innovations: Instruction fine-tuning and reinforcement learning from human feedback (RLHF). Efficient Scaling: Techniques for training and optimizing LLMs at unprecedented scales. 4️⃣ Future-Proof Knowledge: The book outlines new paradigms in AI research, making it an essential resource for researchers, developers, and business leaders. As AI continues to transform industries, understanding the latest advancements in LLMs is critical for staying competitive. #AI #GenerativeAI #LLMs #ArtificialIntelligence #Innovation #TechTrends #AI2025 #FutureOfWork #ML

58

2 Comments

Klarity is a nice library to debug llms decision making process and see how models think → you can measure model confidence using entropy and similarity metrics → works on vlms and you can see how the model attends to images → see how llms reason by evaluating step-by-step thinking patterns → get analysis in a structured JSON format github repo - https://lnkd.in/gNWJRaZJ

108

2 Comments

I want to share this great resource by Sulaiman Shamasna. This 27-minute Medium article, clearly explains how to build an #LLMOps pipeline from data ingestion to model deployment and monitoring. It also contains several visualizations and code snippets which really help in the understanding of the topic. IMHO, great job! 🦾 https://lnkd.in/ddVbZkpS #datascience #ai #machinelearning #deeplearning

55

2 Comments

Key Insights from Kevin Weil (CPO at OpenAI) in conversation with Varun Mavya and Tanmay Bhatt: AI models are "teach-limited" rather than "intelligence-limited." Effective training relies heavily on providing the right data combined with reinforcement learning techniques. Are AI models gonna be commoditized? No. Kevin argues against it saying a short lead of 3-6 months is still valuable because it allows OpenAI to launch new products based on their new capabilities. The cost of intelligence is rapidly declining, becoming roughly ten times cheaper each year. A groundbreaking AI model capable of surpassing human-level performance in competitive coding is expected within the next year. Personalized AI-driven tutoring has substantial potential to transform education, as individualized learning significantly enhances student outcomes compared to traditional methods. Current challenges with AI Voice Mode include premature interruptions during pauses and the inability to replicate nuanced human conversational cues, such as knowing when to interject or yield conversational space.

19

🚀 AI is evolving faster than an IPL run chase! 🏏 AI agents are out here changing the game. Take LangGraph, for example, an open-source platform that’s redefining multi-agentic automation. It lets you create autonomous AI agents that work together seamlessly, making workflows smarter and more efficient. Imagine this: 🤖 One agent fetches emails/chats with critical issue updates. ⚡ Another analyzes and classifies them by using various child agents. 🎯 A third logs tickets automatically in ServiceNow in right category, routes it to correct agent and also sends a level1 response back to the user. 💡 And yet another pulls insights from past resolutions to draft responses. All working together autonomously, just like a well-synced cricket team chasing down a target in the final overs! 🏆 So, with AI agents doing the heavy lifting, we humans can finally focus on the real question: Hyderabadi or Kolkata biryani for the match party? 🍛😆 #AI #Automation #MultiAgentSystems #LLM #Langchain #LangGraph #FutureOfWork #TechInnovation #IPL2025

46

1 Comment

🚀 Kicking off an exploration — from building generic to advanced AI agents, and deploying them in GCP using all native services! All hands-on Starting with basic agent, ever wondered how to implement a basic search agent with Google's ADK? That’s exactly what I built in Part 1 of my new series: AgenticAI with #ADK. In this part, I cover: ✅ What ADK (Agent Development Kit) is and why it’s great for building agents. ✅ Adding the built-in Google_search tool for real-time results. ✅ Testing locally with ADK Web and the ADK CLI. ✅ Debugging and inspecting tool calls like a pro. If you’re into LLMs, tool-using agents, and production-ready AI, you’ll want to follow along. #ADK #AIAgents #GoogleSearchAPI #GeminiAI #FastAPI #Python #GoogleCloud #VertexAI #MachineLearning #LLM #AIEngineering

19

A lot of ml concepts discussed on linkedin are either on a very surface level or sometimes too complex. I thought about writing about things I learn here and there and share my understanding in a no fluff very easy to understand manner. Here is a very brief understanding of Gaussian Mixture Models -> an unsupervised clustering algorithm usually overlooked in clustering algorithm discussions. They work on the assumption that your data belongs to a mixture of gaussians. Read more about it here: https://lnkd.in/gF8EPN5b This is my first time writing. Let me know of any improvements or feedbacks :) #clustering #unsupervisedlearning

35