Projects

Led the development of two 'Suggested Prompts' features in the Acrobat AI Assistant Beta, significantly enhancing user engagement by influencing approximately 20% of customer interactions based on early measurements. Collaborated with cross-functional teams to integrate advanced AI capabilities, improving the overall user experience with intelligent suggestions. This project showcases a successful application of AI to drive user interaction and satisfaction.

Managed the development of two of the team's seven LLM-based editing features, substantially contributing to the product's AI capabilities. Designed and implemented a robust prompt engineering framework, enabling broader use across the organization and streamlining the development of AI features. This framework accelerated the deployment of AI functionalities and promoted consistency in prompt design.

Co-led data acquisition and preparation efforts for a Japanese Document Object Detection model, essential for expanding the product's capabilities into new language markets. Employed weak-labeling techniques to optimize annotation efforts and budget usage, effectively improving data quality while reducing costs. This project expanded the model's applicability and opened up new market opportunities.

Drove an online training initiative by developing a model-agnostic SDK for continuous improvement of AI models. This SDK facilitated ongoing enhancement of model performance by enabling real-time learning from new data. The initiative resulted in significant improvements in key performance metrics, demonstrating the effectiveness of continuous model training in production environments.

Developed and maintained an MLOps service for real-time monitoring of a tables model in production. Provided critical insights into model performance, data drift, and operational metrics, directly informing data and model development efforts. Enhanced the team's ability to respond to issues promptly and maintain high model performance.

Leveraged Dask and big data technologies to process and analyze vast amounts of customer usage data, extracting valuable insights into user behavior and model performance. Initiated early efforts for a tables monitoring service, laying the groundwork for future real-time monitoring and data-driven decision-making. These efforts contributed to a deeper understanding of customer needs and informed subsequent development priorities.

Furthered multimodal representation learning during an internship at Adobe by utilizing triplet networks to enhance model capabilities. Developed models that effectively integrated multiple data modalities, improving the system's ability to understand and process complex inputs. Contributed to advancing the state-of-the-art in multimodal AI applications within the company.

Developed a neural network ensemble using TensorFlow to predict outcomes of English Premier League games, achieving a 73% accuracy rate on matches during the 2017-2018 season. Combined multiple neural network models to improve prediction accuracy and presented the project at Stanford University. Demonstrated the potential of machine learning in sports analytics.

Designed and implemented transmitter and receiver circuits using two Arduinos to send and receive data efficiently. Achieved reliable data transmission in 2,500 microseconds, with potential for faster speeds at reduced reliability (1,200 microseconds). Demonstrated proficiency in hardware programming and real-time communication systems; detailed writeup and code are available upon request.

Built a flashcard application as part of CS193x coursework, implementing features for creating, editing, and reviewing flashcards. Utilized GitHub for version control and collaboration, adhering to best practices in software development. Provided hands-on experience with application development and user interface design.

Developed a music visualizer application that integrates with the GIPHY API to display dynamic visuals corresponding to the music being played. Showcased skills in API integration, real-time data processing, and creative coding. Enhanced the user experience by combining auditory and visual stimuli.

Implemented a custom heap allocator in C, providing functions for malloc, realloc, and free. Achieved 56% memory utilization and 103% throughput compared to standard implementations. Deepened understanding of memory management and low-level programming; code is available upon request.

Designed and constructed a solar-powered USB charger capable of powering any USB device. Involved circuit design, component selection, and assembling a functional prototype. Demonstrated practical application of electrical engineering principles and a commitment to sustainable technology.