Hi, I'm Josh Yeung
Software Engineer at Amazon building scalable systems for Amazon Ads and Prime Video. Passionate about LLMs, web applications, and matcha.
Where I've Worked
Building impactful software at leading tech companies
Amazon
Software Development Engineer II & Machine Learning Engineer
Amazon Ads, Prime Video
Aug 2024 - Present
Seattle, WA
- • Designed and implemented an optimized caching architecture for search query understanding, reducing unnecessary ad sourcing queries with low shopping intent. Resulted in a 16.24% increase in Return on Ad Spend (RoAS).
- • Architected and led end-to-end redesign of visual sourcing multimodal embeddings pipeline, cutting storage costs by 75% and reducing image processing volume by 85.27% (5M to 750K daily), saving $6,600+ monthly.
- • Developed and deployed internal AI assistant based on secure RAG-powered knowledge base leveraging multiple LLMs via Amazon Bedrock, streamlining team workflows.
Software Engineer
Google Assistant
Sep 2022 - Apr 2024
Seattle, WA
- • Designed and launched fallbacks for home/work journeys on Google Assistant (~80M daily users), improving query reliability for 8M users and increasing successful directions queries by 22.68%.
- • Built personalized restaurant recommender using Bard/Gemini, enhancing conversational AI experiences.
- • Developed suggestion buttons for new Google Assistant on mobile, driving 40M daily impressions and 3M DAUs.
- • Integrated 10+ local search features (e.g., "Starbucks hours," "Seattle traffic"), improving navigation capabilities.
- • Designed and implemented "Search Along Route" prototype for Google Assistant (1.5M DAUs).
- • Increased test coverage for local journeys by 400% through automated parameterized regression tests.
Qualtrics
Research Engineer Intern
Qualtrics IQ
May 2021 - Aug 2021
Seattle, WA
- • Improved Qualtrics IQ's text analysis capabilities using NLP, leading to more reliable experimental results.
- • Optimized NLP model deployment on AWS SageMaker and internal platform. Achieved 80% data reduction for Qualtrics dataset inference while maintaining high accuracy (58% F1 score) in aspect opinion tagging.
Academic Background
University of California, Berkeley
Bachelor's degree, Computer Science
Activities and societies: Upsilon Pi Epsilon (UPE)
Things I've Built
Side projects and experiments I've worked on
Pokemon Team Recommender
ML-powered competitive team builder using LightGBM. Trained on 10K synthetic teams with 93% code coverage. Includes validation study comparing synthetic vs. real battle data.
Two Approaches:
Validation Research:
Neural Network Optimizer Robustness Analysis
Research at UC Berkeley RISELab investigating how optimizer choice (SGD, Adam, AdaHessian) affects model robustness to input perturbations. Adam-trained models showed 2-3x better accuracy under realistic noise compared to SGD.
Research conducted at: Berkeley RISELab under Prof. Michael Mahoney and Ben Erichson
Bible Study App
AI-powered cross-platform scripture study app with verse highlighting, note-taking, and chat history management. Built with Capacitor for iOS/Android.
Personalized Restaurant Recommender System
Conversational AI for Google Assistant providing personalized restaurant recommendations based on cuisine preferences, dietary restrictions, and location.
Pokemon Classifier via Resnet
Fine-tuned Resnet34 for Pokemon identification achieving 94% accuracy. Deployed as a public web application on AWS Elastic Beanstalk.
Experimental Bytes
Follow @experimentalbytes where culinary art meets food science! Showcasing creative cooking experiments and delicious recipes.
Pacman AI
UC Berkeley CS188 project implementing search algorithms, multi-agent systems, reinforcement learning, and probabilistic inference in the Pacman domain.
Pokemon Personality Quiz
Find your Pokemon match through 10 fun scenario-based questions. Uses cosine similarity across 24-trait vectors to match you with one of 16 Pokemon personas, complete with personalized results and shareable OG images.
EECS 127 Crash Course for ML/SWE
Comprehensive hands-on crash course covering optimization fundamentals from UC Berkeley's EECS 127. Seven Jupyter notebooks with implementations of SVD, PCA, recommender systems, LASSO, logistic regression, portfolio optimization, and compressed sensing.
Technical Expertise
Technologies and tools I work with across projects
💻 Languages
🎨 Frontend
⚙️ Backend & Databases
🤖 ML & AI
📱 Mobile
☁️ Cloud & DevOps
⚡ Core Skills
Hobbies & Interests
When I'm not coding, you'll find me playing competitive games or enjoying strategic challenges
Table Tennis
Fast-paced rallies and quick reflexes. Played with the Cal (UC Berkeley) table tennis club in college.
Nothing beats a competitive ping pong match!
Tennis
Love the strategy and athleticism of tennis. Favorite racket: Head Radical Pro. Favorite player: Roger Federer.
Great way to stay active and competitive!
Chess
Strategy, tactics, and endless possibilities. Challenge me online!
Board Games
From strategy games to party games, I enjoy the social aspect and creative gameplay.
Current favorites: Wingspan & Chameleon
Places I've Lived 🌍
Based in Seattle, WA — here's a quick globe view of my home cities
Cities
- Seattle, WAHome base (PT)
- Richmond, VACapital of Virginia
- Berkeley, CABay Area vibes, where I went to college at UC Berkeley
- West Lafayette, INCollege town where Purdue is
The Art of Matcha
Beyond being a delicious beverage, matcha is a superfood packed with antioxidants, metabolism-boosting properties, and provides sustained energy without the jitters.
Sustained Energy
Gentle caffeine boost that keeps me focused throughout the day
Rich in L-Theanine
Promotes relaxation and mental clarity without drowsiness
Latest Writings
Thoughts on software engineering, technology, and matcha
Beyond Accuracy: Why Your Optimizer Choice Matters for Real-World ML
Research at UC Berkeley RISELab showing Adam-trained models maintain 2-3x better accuracy under noise
The Art of Brewing the Perfect Cup of Matcha
Discover the traditional techniques and modern twists for making exceptional matcha