LLM Response Quality Classification
A reproducible four-label classification workflow for evaluating accuracy, completeness, reasoning quality, and hallucination risk.
- 150+
- reviewed outputs
- 1.00
- macro F1 demo
- 2
- baseline families
Seattle, WA · U.S. Citizen
Jiangbin Huang
Data Scientist · Applied AI · Analytics Engineering
I build validated data pipelines, statistical and machine learning workflows, and LLM evaluation systems that turn messy inputs into reproducible decisions.
- 1.2M+
- sensor records processed monthly
- 40%
- SQL query performance improvement
- 200+
- manual reporting hours saved annually
- 150+
- LLM responses evaluated
Selected work
Production-oriented projects
Three-Layer Data Lake Pipeline
Configuration-driven ingestion with raw snapshots, validation, rejected-record handling, curated customer metrics, and SQL publishing.
- 3
- data layers
- 98.4%
- valid demo rows
- 3
- reason-coded rejects
Point-in-Time S&P 500 Return Modeling
A leakage-aware equity panel using historical index membership, liquidity controls, walk-forward model comparison, factor regimes, placebo tests, moving-block bootstrap, Reality Check, and locked holdout evaluation. The final analysis reports an exploratory signal rather than claiming persistent alpha.
- 704
- historical members
- 8
- locked OOS quarters
- 0.04
- locked OOS R²
Experience
From data contracts to model review
Co-Founder & Data Scientist
Early-Stage Data & AI Startup
Built Python and SQL workflows, a three-layer data lake pattern, model-selection pipelines, RAG evaluation, and repeatable technical handoff assets.
Data / AI Researcher
Feelie
Standardized 500K+ conversation records and connected DistilBERT emotion classification, LLM response generation, and rubric-based output review.
Data Analyst Intern
UW Engineering Services
Built validated ETL and SQL reporting workflows, improved query performance by 40%, and automated 200+ hours of annual reporting work.
Research Assistant
UW Department of Civil Engineering
Processed 1.2M+ sensor records per month with timestamp alignment, missing-value treatment, noise filtering, and feature engineering; built congestion-monitoring datasets, applied clustering to identify recurring traffic patterns and high-risk locations, and evaluated ARIMA and LSTM forecasting approaches for short-term traffic prediction.
Capabilities
Methods backed by working pipelines
Statistics & Machine Learning
Model selection, feature engineering, cross-validation, hypothesis testing, regression, classification, clustering, and time-series forecasting.
XGBoost · Random Forest · Logistic Regression · PCA · ARIMA · LSTM
Applied AI & LLM Systems
RAG workflows, response-quality classification, retrieval evaluation, prompt testing, embeddings, hallucination review, and model evaluation.
PyTorch · scikit-learn · TensorFlow · LangChain · OpenAI API
Data Engineering & Analytics
ETL development, data validation, layered data lakes, schema design, SQL optimization, analytics datasets, and reporting automation.
Python · SQL · pandas · AWS · Tableau · Power BI
Education
Applied mathematics foundation
Master of Science, Applied Mathematics
University of Washington · Seattle, WA
Scientific computing, machine learning for large-scale data, database systems, probability, and statistical modeling.
Bachelor of Science, Applied Mathematics
University of Washington · Seattle, WA
Data structures, algorithms, numerical analysis, statistical methods, and computational mathematics.