Wei-Hsien (Rachel) Wang — Machine Learning Engineer (LLM Systems & MLOps)

Professional Experience

AI/ML Engineer

Tatung System Technology, Taipei, Taiwan Jul 2025 – Sep 2025

Collaborated with product team to optimize client-facing customer-service workflows, reducing end-to-end latency by 60% through Model Context Protocol (MCP) integrations and Redis caching on Azure
Fine-tuned a RoBERTa intent classifier with confidence-based routing, increasing RAGAS faithfulness by 9%
Engineered RAG–GPT integration using RESTful microservices, raising customer satisfaction (CSAT) by 8 percentage points

Model Context Protocol (MCP) Redis Azure RoBERTa RAG RESTful Microservices Python

AI Engineer

Praxis Solutions, San Diego, CA Mar 2025 – Jun 2025

Built a neuro-symbolic AI system processing ~500K enterprise emails with 87% extraction precision after human-in-the-loop validation using LangGraph on GCP (Cloud Run, Cloud SQL, Vertex AI)
Developed GraphRAG (FAISS + Neo4j) with Gemini, analyzing user query paths and reducing task discovery time by 65%
Coordinated a cross-functional effort to develop a Streamlit dashboard for executives, reducing project delays by 40%

LangGraph GraphRAG Google Cloud Platform Vertex AI Neo4j Gemini Streamlit Neuro-symbolic AI

Research Assistant

Soochow University, Taipei, Taiwan Oct 2023 – Mar 2024

Developed PAD (Prediction And Decision) methodology bridging ML predictions with business strategy implementation
Presented at the 2024 International Conference on Big Data and Enterprise Resource Management
Demonstrated automated marketing strategies and ROI improvement through targeted customer segmentation

R Machine Learning RFM Analysis Public Presentation

Data Analyst Intern

Virbac, Taipei, Taiwan Sep 2023 – Dec 2023

Built SQL ETL pipelines populating a centralized analytics data warehouse from 18 branches for executive ad-hoc reporting
Designed Tableau dashboards illustrating profitability and OPEX/CAPEX trends, enabling ~10% cost optimization
Streamlined reporting workflows with Excel, reducing manual processing time by 32%

SQL Tableau Excel Financial Analysis

Teaching Assistant

Soochow University, Taipei, Taiwan Sep 2022 – Jan 2023

Teaching Assistant for Applications of Data Science course
Led sentiment analysis and text mining projects in R on visitor reviews for the museums in Taiwan
Supported students in data analysis while fostering an inclusive learning environment

R Text Mining Sentiment Analysis

Data Scientist Intern

Growth Strategy, Taipei, Taiwan Jun 2022 – Aug 2022

Analyzed eCommerce history data in R for a cosmetics store (120K+ customers) to identify retention and churn
Applied K-Means clustering and a rule-based New-Exist-Sleep segmentation to classify customer life stages and engagement
Built Logistic Regression (85% accuracy, AUC = 0.88) and Linear Regression (R² = 0.713) models to estimate CLV, and simulated marketing ROI through A/B testing of campaign cost vs uplift in purchase probability

R K-Means Clustering Logistic Regression A/B Testing Predictive Modeling Customer Segmentation

AI Agents

📧

GraphRAG Task Management System

Production AI System @ Praxis Solutions

Production-ready neuro-symbolic AI system that transforms 500K+ enterprise emails into an intelligent, queryable knowledge graph using LangGraph, Neo4j, and Gemini.

                                    
                                    GraphRAG Architecture: Hybrid retrieval system combining
                                        Neo4j knowledge graph with FAISS vector search and OpenAI embeddings for
                                        topic-centered semantic queries
                                

                                    
                                    LangGraph ETL Pipeline: Orchestrated multi-stage workflow
                                        with RAG-enhanced extraction, auto-fix JSON validation, and human-in-the-loop
                                        validation achieving 87% precision
                                

                                    
                                    RAGAS Evaluation: Automated quality assessment with
                                        Faithfulness and Context Recall metrics, providing confidence scores for each
                                        query response
                                

                                    
                                    Dual Storage: Persistent data layer with PostgreSQL (Neon)
                                        and Neo4j graph database storing entities (Person, Task, Topic, Date) and
                                        relationships (ASSIGNED, RESPONSIBLE_TO, COLLABORATED_BY)
                                

                                    
                                    Natural Language Queries: LLM-powered Cypher generation
                                        enabling questions like "What tasks did Alice assign to Bob due this week?" with
                                        interactive graph visualization
                                

                                    
                                    Reduced task discovery time by 65% and project delays by
                                        40% through intelligent email parsing and deadline
                                        tracking
                                

Tech Stack: LangGraph, Neo4j, PostgreSQL (Neon), OpenAI GPT-4o, FAISS, RAGAS, Streamlit, Docker, GCP (Cloud Run, Vertex AI)
Key Features: Email metadata extraction, LLM task extraction, human-in-the-loop validation, knowledge graph construction, semantic search, interactive visualization, LangSmith tracing

LangGraph GraphRAG Neo4j PostgreSQL GPT-4o FAISS RAGAS Streamlit Docker GCP

🎧

Muse.AI: The Agentic Music Curator

Gemini 3 Hackathon

Agentic music curation system built with Gemini 3 Flash and LangGraph. Agents collaborate to analyze visual vibes from photos and curate personalized playlists with critic-validated reasoning.

                                    
                                    Multimodal Agentic Workflow: Orchestrated Vision, Search, and
                                        Curator agents using LangGraph for multi-step reasoning and
                                        self-correction
                                

                                    
                                    Visual Vibe Analysis: Utilization of Gemini 3
                                            Flash to extract abstract mood and narrative from images,
                                        translating pixels into audio parameters
                                

                                    
                                    The Critic (Quality Control): Specialized agent that
                                        validates recommendations against context, preventing hallucinations and
                                        ensuring high relevance
                                

                                    
                                    Long-Term Memory: Firestore-backed user profile that learns
                                        preferences over time, biasing future recommendations while respecting current
                                        context
                                

Competition: Gemini 3 Hackathon, Google DeepMind

Gemini 3 Flash LangGraph Spotify API Next.js FastAPI Firestore Google Cloud Run

🍽️

Recipe Recommendation System

Web Mining & Recommender Systems

Applied machine learning focused on recommendation systems and text mining with collaborative and content-based filtering approaches.

Collaborative filtering with SVD

Content-based filtering with TF-IDF

Steam reviews classification

79% accuracy on multi-cuisine prediction

Team: Darren Trieu, Wei-Hsien (Rachel) Wang, Akshay Ghosh, Mathew Luo

Collaborative Filtering SVD TF-IDF XGBoost Content-Based Filtering

🌍

AI Recruitment Hub

Deep Learning & GenAI for Business

Full-stack AI Recruitment System automating the hiring lifecycle from resume parsing to offer letter generation. Integrated LLaMA and Gemini APIs for candidate assessment and contract generation.

Automated resume parsing & analysis

Calendly integration for scheduling

Dynamic PDF generation

Containerized deployment

Team: Ryosuke Aoguchi, Nicole Ou, Wei-Hsien (Rachel) Wang

LLaMA Gemini API Python Shiny Docker PDF Generation

Modeling & Experiments

📂

Customer Analytics Cases

Customer Analytics

Comprehensive business analytics covering predictive modeling, A/B testing, and profit optimization across real-world customer scenarios.

Tuango: Mobile app message targeting with logistic regression

TZ Gaming: Optimal mobile ad targeting with ROME analysis

Intuit: QuickBooks upgrade prediction using neural networks

Creative Gaming: Propensity-to-buy & uplift modeling

Pentathlon: Next product recommendation system

S-Mobile: Customer churn prediction and retention strategy

FiTech: Customer adoption and value forecasting

Team: Ryosuke Aoguchi, Wei-Hsien (Rachel) Wang, Jaqueline Vallejo, Qiyong Zhang

Logistic Regression XGBoost Neural Networks Random Forest A/B Testing Uplift Modeling

🎼

Generative AI & Audio Processing

Machine Learning for Music

Complete progression through modern music AI, from audio signal processing to symbolic and generative modeling using state-of-the-art techniques.

MAESTRO 2018 dataset for piano performances

Markov chains + LSTM for melody generation

REMI tokenization & Transformer decoder

Diffusion models for audio generation

Task 1: Unconditioned Symbolic Generation

Separate harmony and melody generation using Markov chains and LSTM
Temporal and harmonic alignment evaluation

Task 2: Conditioned Symbolic Generation

Multi-instrument music with REMI tokenization
Structured generation with intro, climax, and resolution
Expressive decoding with MLP regressors

LSTM Transformer Markov Chains REMI Diffusion Models Audio Processing

🎆

NLP & Deep Learning

Analyzing Unstructured Data

Comprehensive exploration of NLP and deep learning for unstructured data, covering traditional text classification, embeddings, and prototype-based learning.

Sentiment Classification: TF-IDF + logistic regression on Financial Phrase Bank

Word Embeddings: Word2Vec vs. GloVe comparison and n-gram modeling

Phrase Mining: AutoPhrase on CS papers with domain-specific embeddings

Restaurant Classification: Multi-class classification with 81% F1 score

Image Prototyping: MNIST classification via K-Means + KNN vs CNNs

Final Project Team: Ryosuke Aoguchi, Nicole Ou, Wei-Hsien (Rachel) Wang

TF-IDF Word2Vec GloVe AutoPhrase CNN K-Means

Data & Cloud Engineering

🔁

AWS Cloud Analytics

Scalable Analytics

Comprehensive cloud-based analytics using PySpark, AWS EMR, and SageMaker. Projects include large-scale text processing, airline delay analysis, and salary prediction.

AWS EMR for distributed computing

SageMaker for ML deployment

Performance benchmarking

Production-level workflows

PySpark AWS EMR SageMaker XGBoost Distributed Computing

⚡

Distributed Systems Engine with Ray

Scalable Data Systems

Built production-grade distributed data processing systems using Ray for parallel computing and GPT for intelligent data transformation.

                                    
                                    Built ETL pipeline with Ray Data API and
                                        GPT to process 100K+ Amazon reviews via
                                        streaming execution
                                

                                    
                                    Implemented MapReduce to distribute word-count processing
                                        across eight parallel workers, validating 2.8K+ unique
                                        terms
                                

                                    
                                    Engineered and benchmarked three AllReduce algorithms
                                        (recursive-doubling BDE, binary-tree MST, Ray builtin) using
                                        PyTorch tensors, profiling 257 MB/s bandwidth
                                        with <1% variance
                                

Course: Scalable Data Systems, Halıcıoğlu Data Science Institute

Ray MapReduce AllReduce PyTorch GPT Distributed Computing ETL

📦

Cloud Data Warehouse with Snowflake

SQL & ETL

Advanced SQL and ETL workflows across PostgreSQL and Snowflake. Integrated multi-format datasets with automated ingestion and materialized views for business intelligence.

PostgreSQL & Snowflake integration

Recursive CTEs & fuzzy matching

Cloud-based materialized views

BI pipeline design

Final Project: Automated ingestion of CSV, XML, PostgreSQL, and Marketplace APIs into Snowflake using Python. Computed PO amounts, invoice discrepancies, and joined purchasing and weather data for analytics on price variance under different temperature conditions.

Team: Yangqi Huang, Wei-Hsien (Rachel) Wang, Zewei Chen, Guoxuan Li

PostgreSQL Snowflake Python ETL SQL Data Pipelines

Education

🎓

Master of Science in Business Analytics

Rady School of Management, UC San Diego

Aug 2024 - Dec 2025

Relevant Courses:

Customer Analytics Scalable Data Systems SQL & ETL Web Mining & Recommender Systems Deep Learning & GenAI Machine Learning for Music Analyzing Unstructured Data

🏛️

Bachelor of Business Administration, International Business

Soochow University, Taipei, Taiwan

Sep 2020 - Jun 2024

Certificates & Achievements

AWS Certified Cloud Practitioner

Amazon Web Services

Introducing Generative AI with AWS

Amazon Web Services

Microsoft Azure AI Engineer Associate (AI-102)

Microsoft

SAS Cortex Participant

SAS Institute

IBM Deep Learning Essentials

IBM

Conference Presentation

2024 International Conference on Big Data and Enterprise Resource Management