Leo YaoResume ↗

Hi, I'm Leo.

AI Systems Engineer
& Technical Leader

I’m Leo, a Founding Engineer at Agaruda, building an enterprise AI platform and leading engineering execution across five teams and 15 people.

My experience spans AI systems, cloud infrastructure, SaaS, and automotive software across Agaruda, Proto, Dell, and Garmin, combining hands-on engineering with cross-team technical leadership.

I hold an M.S. in Computer Science from National Taiwan University and was a Visiting Researcher at Tsinghua University. My research focuses on medical AI, clinical reasoning, and reinforcement learning, with seven publications, 200+ citations, and one published US patent application.

Resume ↗LinkedIn ↗Google Scholar ↗
Leo Yao
Leo Yao

Expertise

The areas I work across.

AI Systems

AI agents · Multi-agent orchestration · RAG · MCP · Human-in-the-loop

vLLM, LiteLLM, Agno, pgvector, PyTorch

Backend & Cloud

Multi-tenancy · APIs · Authorization · Observability · Hybrid deployments

FastAPI, PostgreSQL, Redis, AWS, Kubernetes/RKE2

Software Engineering

System design · Production debugging · Automation · Developer tooling

Python, Java, TypeScript, React, Linux, CI/CD

Technical Leadership

Cross-team communication · Requirements · Delivery · Feature ownership

Architecture reviews, Scrum, Sprint Master, Team Lead

Professional Experience

Resume ↗
Agaruda conceptual system flow
AgarudaJun 2025 – Present

Enterprise AI & data center platform ↗

Founding Engineer · Senior Software Engineer & TPM

Led execution across five teams and 15 people, connecting Backend/Infra, Frontend, AI, 3D/Omniverse, and Design.

My contribution

Technical decision Designed multi-agent orchestration, tenant-aware state, and hybrid deployment architecture.

Outcome Built platform capabilities across 10 core services and 50+ infrastructure components.

LeadershipAligned requirements, dependencies, and communication across five teams and 15 people, working with a core three-engineer Backend/Infra team.

FastAPI · Kubernetes/DevOps · AI Agent · vLLM · On-Prem Deployment

Proto conceptual system flow
Proto.cxAug 2024 – Jun 2025

AI customer experience SaaS ↗

Full Stack Engineer

Built AI and full-stack features across customer conversations, analytics, billing, and integrations.

My contribution

Technical decision Optimized analytics queries and developed AI usage tracking for a multi-tenant platform.

Outcome Delivered 15+ features; reduced analytics API database round trips by approximately 80%.

React · FastAPI · PostgreSQL · RAG · LLM Integration

Dell conceptual system flow
DellMay 2022 – Mar 2024

VxRail & cloud lifecycle management ↗

Software Engineer II

Worked on multi-cluster management and upgrade reliability for hybrid VMware/AWS infrastructure.

My contribution

Technical decision Improved REST APIs and storage workflows, with root-cause analysis of customer-reported issues.

Outcome Improved upgrade success to 84%; resolved 50+ customer issues.

Spring Boot · Django · Go · RKE2

Garmin conceptual system flow
GarminJun 2021 – Apr 2022

Automotive software & diagnostics ↗

Software Engineer

Developed diagnostic and logging tools for Toyota in-vehicle systems.

My contribution

Technical decision Built a multithreaded Linux diagnostic application and searchable Java/Shell logging service.

Outcome Monitored 23 hardware components; reduced factory troubleshooting time by 20%.

Android Car · Linux · Java · Shell · Hardware integration

Independent & Collaborative Projects

Products and collaborations beyond my primary roles.

約美 YueMe product preview
約美 YueMeJul 2026 - Present

Beauty booking & member management ↗

Product Building

A product for beauty-service appointments and member management.

My contribution

Leading a small R&D team of three, including market research and promotion, and quickly building and deploying the system from scratch.

Full Stack · Go · DB design · Marketing

XChange Xtalent public website illustration
XChange XtalentJul 2022 – Present

Mentorship & talent platform ↗

Team Lead & Backend Engineer

Directly built/led a cross-countries team based on TW, Canada, the US to deliver cloud-based mentorship services.

My contribution

Technical decision Coordinated backend development, Scrum execution, and cross-functional delivery.

Outcome Delivered two customized SaaS platforms within budget and schedule.

LeadershipDirectly led a 10-member distributed team across Taiwan, Canada, and the US, coordinating Scrum execution and delivery.

AWS · Backend architecture · Scrum · Project Management

TaiDing Biotechnology Inc. project workflow illustration
TaiDing Biotechnology Inc.Jul 2021 – Present

Livestock AI & genetic management platform

Founding Engineer · Full-Stack & Machine Learning

Built and deployed livestock AI and a cloud-based genetic management platform for two leading pig breeding farms.

My contribution

Applied AI Developed a YOLO- and VideoMAE-based real-time farrowing detection pipeline, achieving 75%+ accuracy.

Full-stack development Built genetic data management, inbreeding coefficient analysis, and breeding optimization services from scratch with React, FastAPI, PostgreSQL, and AWS.

Data pipeline Used Airflow and Kinesis to support the detection pipeline.

PyTorch · YOLO · VideoMAE · React · FastAPI · PostgreSQL · AWS · Airflow · Kinesis

Bank of Marshall Islands — BOMI project workflow illustration
Bank of Marshall Islands — BOMIFeb 2023 – Mar 2025

Mobile banking Android application

Project Management & Backend / Android Collaboration

Helped deliver a launched mobile banking app through market research, project management, and coordinated backend and Android development.

My contribution

Product & delivery Conducted early-stage market and requirements research, managed project execution, and coordinated backend and App development.

Mobile engineering Contributed to banking workflows spanning account inquiries, transfers, payments, and loan services.

Application architecture Worked with a layered Kotlin architecture using ViewModels, use cases, repositories, and reusable XML / Jetpack Compose interfaces.

Kotlin · Android · Jetpack Compose · Coroutines · Retrofit

Research & Patents

Medical AI, clinical reasoning, speech analysis, and intelligent systems.

Google Scholar ↗
7Papers
200+Citations
1US patent application
Earlier RPRO preprint workflow · published RGPO figure pendingEarlier RPRO preprint workflow · published RGPO figure pending

2026 · IEEE Transactions on Artificial Intelligence

RGPO: Ranking-Guided Preference Optimization for Reliable Clinical Reasoning ↗

Abstract summary

RGPO refines clinical reasoning through task-adaptive templates and ranking-guided preference optimization. Groupwise ranking with the Bradley–Terry model and KL regularization guides learning. Evaluation includes PubMedQA, MedQA-USMLE, and a clinical dataset from Far Eastern Memorial Hospital.

Read paper ↗
Original Figure 1 · Yao et al., JMIR · CC BY 4.0Original Figure 1 · Yao et al., JMIR · CC BY 4.0

2021 · Journal of Medical Internet Research

A Novel Deep Learning–Based System for Triage in the Emergency Department Using Electronic Medical Records: Retrospective Cohort Study ↗

Abstract summary

This study combines structured clinical information and free-text records to estimate hospital admission needs. A text-based representation feeds complementary convolutional and recurrent models. The combined approach was evaluated using US emergency-visit records and data from National Taiwan University Hospital.

Read paper ↗
Original Figure 1 · Tsai et al., Scientific Reports · CC BY 4.0Original Figure 1 · Tsai et al., Scientific Reports · CC BY 4.0

2021 · Scientific Reports

An efficient context-aware screening system for Alzheimer's disease based on neuropsychology test ↗

Abstract summary

A neural screening approach analyzes speech transcripts from neuropsychological tests to distinguish cognitive groups. Attention-based representations reduce manual preprocessing and feature selection. Experiments on English and Chinese datasets examine classification of Alzheimer's disease and mild cognitive impairment against healthy controls.

Read paper ↗
Simplified system flow redrawn from the abstractSimplified system flow redrawn from the abstract

2020 · IEEE International Conference on Systems, Man, and Cybernetics

A System for Predicting Hospital Admission at Emergency Department Based on Electronic Health Record Using Convolution Neural Network ↗

Abstract summary

Emergency records are transformed into image representations for a convolutional classifier that predicts hospital admission. The approach was evaluated on 118,602 US emergency department visits, reporting an AUROC of 0.86 and accuracy of 0.77.

Read paper ↗
Topic illustration · original architecture figure pendingTopic illustration · original architecture figure pending

2020 · International Conference on Smart Grids and Energy Systems

A Demand Side Management with Appliance Controllability Analysis in Smart Home ↗

Abstract

As global energy consumption keeps increasing, the energy efficiency has become an important issue. Demand-side management (DSM) system is a technology that can be used to solve this issue. In many houses, there are some time periods in a day when inhabitants have high electricity demands, and these peaks will cause a burden on the smart grid. In order to solve energy efficiency problem, the demand-side management system can help users to postpone non-immediately needed demands to off-peak hour subject to the mechanism of real-time pricing policy. In addition, with the development of Internet of Things technology, the devices which have internet module can form a home area network, enabling DSM to monitor the entire state of the house and to control appliances. However, due to the uncertainty of user behavior, it is hard to find the good control policy for energy saving, considering that different users may have various habits and the appliance usage tends to be irregular. Therefore, the ability of DSM will be restricted in order to avoid disturbing the user comfort. In the proposed work, a smart home demand-side management system within home area network was implemented, including multiple agents with reinforcement learning for managing different appliances, energy storage system, and renewable energy. The proposed system also keeps the user's comfort based on the activity of user and controllability of appliances. As a result, the simulation studies show that the proposed DSM system can reduce 42.36% peak demand value and achieve 32.79% cost reduction. The cost reduction value can be further increased by 22.41% when the concept of energy self-awareness is combined into demand-side management system.

Read paper ↗
Original Figure 1 · Chien et al., Scientific Reports · CC BY 4.0Original Figure 1 · Chien et al., Scientific Reports · CC BY 4.0

2019 · Scientific Reports

An Automatic Assessment System for Alzheimer’s Disease Based on Speech Using Feature Sequence Generator and Recurrent Neural Network ↗

Abstract summary

Speech is converted into a learned feature sequence and classified by a recurrent neural network for Alzheimer's assessment. The system automates feature generation instead of relying on hand-selected speech statistics. Evaluation on 120 speech samples reports an AUROC of 0.838.

Read paper ↗
Topic illustration · original architecture figure pendingTopic illustration · original architecture figure pending

2019 · IEEE International Conference on Systems, Man, and Cybernetics

A Novel Screening System for Alzheimer's Disease Based on Speech Transcripts Using Neural Network ↗

Abstract

Alzheimer's disease has become one of the biggest challenges in the healthcare system worldwide. Researches have shown that Alzheimer's disease is the sixth leading cause of death in the United States and even the fifth leading cause among people aged 65 and older. Moreover, the number of patients is escalating rapidly in recent years, which also increases the burden on the healthcare system. Therefore, a screening system that can help the doctor to diagnose Alzheimer's disease is demanded. In this paper, we proposed a screening system based on the transcripts of speeches spoken by subjects undertaking a neuropsychology test. While most of the related studies have utilized extracted syntactic and semantic features and relied on a feature selection process, the proposed system used word vectors as the representation of a spoken speech, and Recurrent Neural Network together with attention mechanism as the classifier. Using ten times 10-fold cross validation on an open dataset with 242 speeches samples spoken by healthy controls and 257 samples spoken by subjects with Alzheimer's disease, a mean accuracy of 0.835 is achieved in our work, which is outperforming the current state-of-the-art while requiring less effort.

Read paper ↗
US PATENT APPLICATION

Dynamic timeout with real-time metrics ↗

Adapting operation timeout values to real-time system conditions instead of relying on fixed thresholds.

US20260056816A1 · Published February 26, 2026 · Pending

Inventors: ChihHao Hsu, Guan Liang Chen, LiHung Yao · Assignee: Dell Products LP

Contact

For engineering opportunities,technical conversations, or collaboration.

leoyao816@gmail.com ↗LinkedIn ↗