Huimin Zhang

📖 Exploring Non-Stationary User Behavior in Recommendation Systems Continuing my dive into ByteDance's "Monolith: Real-Time Recommendation System with Collisionless Embedding Table", I’ve come across a crucial challenge: non-stationary distribution—also known as Concept Drift. Unlike visual or linguistic data, which changes over centuries, a user's interests can shift in just minutes. This dynamic behavior makes it difficult for recommendation systems to keep up. 🧠 Why It Matters: Traditional systems struggle because they don’t update fast enough to reflect these real-time changes in user behavior. Monolith addresses this by: Collisionless Embedding: A smarter design that avoids data collisions using a dynamic hash table, improving the use of sparse features. Real-Time Feedback: The system loops user interactions back into training immediately, allowing the model to continuously update and adapt. 🚀 Results: Monolith not only reduces memory usage but also boosts performance (state-of-the-art AUC) without overwhelming server resources. This real-time adaptability is key for delivering more personalized, accurate recommendations! From next post, we will start discussing the design of this system. Keep learning, Keep growing ! Image source : https://lnkd.in/g42JiYKY #AI #RecommendationSystems #ConceptDrift #MachineLearning #DataScience #RealTimeAI #TechInnovation #Monolith #ByteDance #Scalability #LearningJourney

3

Over a Decade of Moonshots: From Search to AI Last Friday marked the end of my 13-year journey at Google, a journey that has transformed me as an engineer, leader, and individual, a journey that I am incredibly grateful for! From my beginnings as a fresh grad to impacting billions of lives through contributions to some of Google's most groundbreaking projects. Beginnings with Search Infrastructure Despite the immense challenges and Infra complexity, I was drawn to an ambitious new moonshot of revolutionizing personal content understanding to improve users’ personal journeys. One of my fond achievements was developing the foundation for private instant indexing, which dramatically reduced the indexing overhead and enabled real-time updates. Revolutionizing Search with Question Answering Next, I joined another early Search Moonshot in Question Answering, a groundbreaking effort to revolutionize search experiences beyond the traditional 10-blue links by fulfilling users' insatiable curiosity for knowledge with concise answers. We pushed the boundaries of QA to what seems like unimaginable frontiers (coverage and quality). I am particularly proud of the progress my team, collaborators, and I made on the Short Answers, Knowledge Augmentation, and overall Answer Freshness to identify trusted answers from an increasingly opinionated web and deliver fresh responses to users (a foundation that became especially critical later, during the Covid pandemic). New era with LLMs In 2019, I felt that we may have reached a performance plateau with our QA/AI strategies, and eyeing the larger Search-wide impact, I embarked on my own moonshot (named MUM) to harness the power of emerging LLM technology to address Search users' needs holistically. I recruited and led a team of believers from Search and Research to develop a groundbreaking internal LLMs enabling a significant uplift to Search experiences and quality, marking a new ML era in Search that was recognized with Google's prestigious 2023 Technical Impact Award. Building on MUM success, I joined the Gemini effort, leading internal Gemini x Search efforts and delivering state-of-the-art novel internal variants aligned with Search needs. A Heartfelt Thank You It truly takes a village and my success wouldn’t have been possible without an exceptional team and peers that joined me in this exciting and fruitful journey! Reminded by a note I once read, "You don’t create a culture. You define values, find like-minded people who live those values, and together a culture is born." I am incredibly proud of the culture we nurtured and humbled by the warm, heartfelt, and incredibly touching goodbyes. Thank you outstanding team, exceptional collaborators, beloved leaders, and Google! I will always cherish the memories and friendships I made and take great pride in this beautiful chapter in my career!

647

47 Comments

Excited to be back in the Bay Area this week to represent Caper and Instacart at the Embedded Vision Summit! Ever curious how to deploy and optimize AI models running on edge devices? I’ll break down some tips and techniques in my talk, "Real-Time Retail Product Classification on Android Devices inside the Caper AI Cart” If you're into AI, machine learning, or just want to learn about the Caper Cart or how we're improving our user's experience by leveraging AI, please feel free to come say hi!

63

2 Comments

🚀 How to Approach a System Design Interview: Amazon Prime Video Example When tackling a system design problem, especially for a complex platform like Amazon Prime Video, it's crucial to approach it strategically. Here's how I break it down: Get clarity on the requirements, both functional and non-functional. Focus on the user flows, define the scale, detail out the components, improve scalability, and wrap up confidently. Pro Tip for Interviews: Don’t rush into drawing diagrams! Ask clarifying questions, understand the problem fully, and then build the system piece by piece. Communicate your thought process clearly, and focus on scalability, reliability, and cost-efficiency. Remember, the trick is to think aloud—explain why you're making certain decisions. The interviewer wants to understand your reasoning as much as the final solution! Good luck to all those preparing for system design interviews! You've got this!💪 Want to master system design interviews and land those $500k+ TPM roles? Make sure you're fully prepared with the Kracd.com interview prep course. Sign up today and transform your career in just 30 days. #SystemDesign #AmazonPrimeVideo #TechInterview #InterviewTips #Scalability #CDN #TechCareers #SoftwareEngineering #Microservices #TechLeadership

9

I'm thrilled to announce that my friend, Melody(Yifan) Gao, and I secured second place🥈 in UChicago Data Science Institute's Gen-AI Hackathon 2024! 📊 Our Technology: 1. Developed a fine-tuned Llama3 model that generates SQL in real-time based on user queries and enhances summarization for question answering. 2. Integrated RAG to store semi-structured data in vectorDB and SQL (PostgreSQL DB). 3. Converted user queries to SQL in real-time with our model, retrieved contextual information quickly from RAG, and used our fine-tuned model to process the data and generate comprehensive answers. 📝 Our Objective: To assist medical providers by facilitating real-time medical insights, streamlining medical report generation, predicting medical conditions, and enhancing decision-making in healthcare. Our solution is 100% open source and 100% original; we did not use any paid models like GPT. We will continue to improve our solution, so please stay tuned! Please visit our GitHub: https://lnkd.in/gtPpBWZJ

77

15 Comments