Engineering the Future of
Agentic AI & Large Language Models
Princeton-trained AI Scientist building production-grade Generative AI systems for finance and accounting. Inventor of 8 US Patents spanning LLM architectures, GNNs, and computer vision. Author of 10+ influential technical articles on agentic AI and LLM fine-tuning.
I'm a Principal Machine Learning Scientist at Sage, where I lead the architecture and deployment of Generative AI applications for finance and accounting. With a Ph.D. from Princeton University and a B.S. from National Taiwan University, I bring rigorous scientific discipline to practical AI systems that matter.
My work centers on Agentic AI — building systems that can reason, plan, and act autonomously. I design multi-agent orchestration pipelines using MCP, Pydantic AI, and LangChain, run distributed LLM fine-tuning with DeepSpeed & Ray on AWS, and ship production ML systems at scale.
I am an inventor on 8 US Patents and Applications spanning Generative AI, LLM architectures, Graph Neural Networks, and Computer Vision — and I actively contribute to the AI community through 10+ technical publications reaching thousands of practitioners globally.
Beyond technical work, I apply AI for social good: using Generative AI to advance math education and create impactful parenting resources for low-resource families.
Inventor on 8 US Patents and Applications spanning Generative AI, LLM architectures, Graph Neural Networks, Computer Vision, and MLOps. Click any patent to view the full filing.
Computer vision framework applying image masking and trained neural networks for accurate product identification and segmentation — enabling scalable visual understanding in multi-modal AI pipelines.
Iterative LLM architectures for handling complex, multi-step queries by dynamically leveraging external APIs and databases. Advances conversational AI beyond single-turn responses into persistent, reasoning-capable agents.
Novel methods for automated, context-aware prompt generation for large language models — improving response quality, task alignment, and downstream performance in enterprise AI applications.
Systematic framework for detecting and mitigating hallucinations in LLM outputs — a critical reliability layer for deploying trustworthy generative AI in high-stakes financial and enterprise environments.
GNN-based framework for constructing and training graph models over complex entity relationships — powering fraud detection, anomaly surfacing, and relational reasoning at enterprise scale.
Methods for detecting covariate drift in production ML systems — enabling continuous monitoring of input distribution shifts to maintain model accuracy and reliability in live financial AI deployments.
Automated methods for identifying relevant data resources and computing derived metrics — providing intelligent data discovery and quantitative insight generation for AI-driven financial analytics.
Adaptive UI personalization system that dynamically tailors interface elements to individual user behavior and preferences — improving engagement and usability in enterprise software products.
Specialized in computational physics and numerical simulation. Developed deep expertise in high-performance computing and mathematical modeling — a rigorous foundation that now informs how I approach large-scale distributed ML systems.
10+ highly cited articles on Medium (Sage AI, Data Science Collective) and Towards Data Science, focusing on LLM fine-tuning, agentic workflows, and scaling laws.
End-to-end implementation of an LLM agent using Model Context Protocol hosts and servers. Integrates open-source and proprietary LLMs via OpenAI-compatible APIs — a practical guide to production-ready agentic architecture.
Pioneering guide on instruction-following fine-tuning for LLMs in a distributed cluster framework. Covers DDP, memory optimization, and acceleration for 3B+ parameter models — a first-of-its-kind resource for the community.
Low-code workflow architecture for building production LLM agents with multi-step reasoning, tool use, and open-source model serving.
Technical bridge between generative and discriminative AI paradigms — adapting LLM next-token prediction for classification tasks.
Deploying a fully local, privacy-first voice assistant using lightweight LLMs — no GPU required. Part of the lightweight LLMs guide series.
Rigorous empirical evaluation of GPT-family models as embedders, with surprising findings about the effectiveness of different embedding strategies.
Multi-disciplinary hackathons organized by Sage Foundation, applying agentic AI to real-world challenges for underserved communities.
Problem: ParentText — a WhatsApp/SMS parenting course reaching families in South Africa, Mexico & Malaysia — saw a 50% completion drop when delivered without human coaches. Static comic strips caused significant drop-off.
Problem: STACK, the world's leading open-source math assessment system, lacked personalized learning paths. Educators spent excessive time compiling data; students had no adaptive progression.
Sharing expertise in Generative AI with the next generation of developers and researchers.
A hands-on course teaching students how to harness Generative AI to build real-world applications. Covers core GenAI concepts, prompt engineering, and practical implementation — empowering the next generation of AI practitioners through project-based learning.
Open to conversations about Agentic AI research, LLM systems, and impactful AI applications. Reach out on any platform below.