Experience

Returning to Google X to continue work on LLM graph orchestration, loop engineering, and code-review agents.

Automated agent development with LLM graph orchestration and loop engineering, producing prototypes with novel capabilities that matched or exceeded those of human-designed pipelines.

Built an LLM code-review agent that replaced the team's internal agent with stronger review capabilities (LangChain, Vertex AI, RAG, agent skills, Google Cloud Platform (GCP), GitHub CLI, Linear).

Researching personalized speech enhancement and target speech extraction for multi-speaker conferencing systems.

Investigating small, real-time-capable generative models for conferencing audio (flow matching, MeanFlow).

Developed an in-context learning framework with large audio language models, achieving SOTA deepfake detection on in-the-wild speech (+54.9% avg Macro F1 improvement) (Natural Language Processing, RAG, ICL, Audio LLMs, and model routing).

Developed an AI tutoring model with 6× higher F1 (AAAI 2025) and LadderSym with 2× better missed-note detection (ICLR 2026) (PyTorch, NLP, Audio Transformers).

Leveraged HPC clusters and Linux environments to engineer generative pipelines for music mistakes, producing two datasets.

Organized 'AI for Music' workshops at AAAI 2025, IEEE ICME 2025, and NeurIPS 2025, showcasing leadership and communication.

Made an app for menu translation, currency conversion, and allergen alerts using Flutter and Android Studio.

Built an MVP in 4 months with weekly sprints, emphasizing leadership and collaboration.

Integrated Google ML Toolkit for cross-platform AI with computer vision and real-time deployment.

Code

Raised cancer-screening invitation accuracy to 83% using scikit-learn and tree-based classification (machine learning, statistics, data science).

Elevated flash-chip yield by 17% with ensembling and Gaussian processes (NumPy, pandas).