Xiangrui (Jerry) Zhao

Developed and maintained software validation tools for complex Process Design Kits (PDKs) in Cadence Virtuoso using C++ and SKILL, enhancing IC design processes by ensuring high-quality PDK validation and migration through robust, integrated solutions within the design environment.

Enhanced tool performance and customer experience by creating user-friendly UI features, streamlining batch operations with foundational APIs, and implementing robust testing frameworks; resolved critical customer-reported issues, increasing efficiency and reducing errors.

Leveraged AI, machine learning, and data-driven methods to generalize PDK validation systems and automate PCell code generation, resulting in reduced test patterns and faster quality analysis without compromising overall quality.

Collaborated closely with customers to implement product enhancements and fix development bugs; aligned system designs with customer needs, improved software quality and maintainability, and addressed challenges in PDK migration and readiness across various Cadence Virtuoso products.

Developed high-fidelity text-to-3D object generation models using GANs and Stable Diffusion. Implemented and optimized state-of-the-art diffusion-based neural network to generate accurate, textured 3D objects from text and images.

Integrated advanced techniques from Stable-Dreamfusion, ControlNet, OpenAI's Point-E, and Pixel2Mesh models to enhance performance and output quality.

Optimized training efficiency and model convergence through hyperparameter tuning using custom accuracy metrics. Resolved instability issues by employing prompt engineering techniques with ChatGPT to design effective text prompts, leading to faster inference times and more accurate results.

Developed a web interface for 3D object generation using AWS EC2 and integrated cloud storage with S3. Deployed models on AWS SageMaker using Docker containers and custom training scripts, automating resource allocation and training pipelines, which streamlined the workflow and improved scalability.

Academic Accolades: 2:1

Core courses: Reinforcement Learning, Deep Learning, Natural Language Processing, Computer Vision and Robotics, Applied Machine Learning, Blockchain and Distributed Ledgers.

Developed a framework for the control of self-organizing robots based on multi-layer neural network architecture, offering better controllability, scalability, and systematic exploration and self-evaluation.

Proposed algorithms and evaluation metrics to balance exploration and exploitation in self-organizing robots, which is interpretable to humans in exploring plausible robot behaviours at a high level.

Proposed the DIAMOND model using deep recurrent neural networks to model information flow among neural assemblies in the brain, achieving guided self-organization in a top-down manner.

Implemented autonomous learning algorithms within the deep homeokinesis model, simulating creativity and curiosity factors for generating novel self-exploratory behaviours for self-organizing robots.

Developed a multi-agent deep reinforcement learning model in the CARLA simulator to train agents in cooperative behaviors, reducing collisions and self-interested behaviors compared to single-agent baselines.

Developed a sample-efficient IRL algorithm based on maximum entropy, allowing agents to learn from fewer expert demonstrations.

Developed a web-based search engine for AI research papers using JavaScript and PostgreSQL, with NLP algorithms for keyword extraction, grammar correction, and query completion prediction.

Developed a decentralized chess game leveraging Solidity smart contracts, ensuring a tamper-proof and transparent game experience.

Developed Android and iOS apps for classifying live human activities using on-device machine learning and IoT sensors.