About
Hi, I’m Hantian. I’m finishing an M.A. in Statistics at Columbia, on the machine learning track, after studying applied math and computational sciences at Duke Kunshan and Duke.
To me, building with AI means putting a model where it actually helps, keeping simple rules where they do the job better, and testing the output before trusting it. A good-looking demo is where the work starts, not where it ends.
Experience
Jun – Aug 2026
Data Science & Machine Learning Intern · Tradingbase AI
- Strategy backtesting
- Out-of-sample validation
- FOMC text features
- AI product research
- Tableau
Feb – Mar 2026
Data Science & Machine Learning Intern · Innovation AI
- Sector rotation
- Random Forest
- XGBoost
- Time-based validation
- SQLite
May – Jun 2024
Strategy Intern · Hua & Hua Marketing Consulting
- Competitive analysis
- Discrete choice models
- Price elasticity
Jun – Jul 2023
Business Intern · China Securities
- Factor models
- Panel regression
- Policy analysis
Projects
Case studies
Evidence Lab
An AI-assisted workspace for reviewing documents, tracing claims to sources, and deciding what the evidence supports.
- AI product research
- Evaluation design
- Python
Market Agent
A market dashboard that ranks what needs attention with plain rules, uses Claude only to explain, and holds the AI-written text to an eval suite.
- 35→78%
- held-out eval pass rate, daily brief
- 75→94%
- held-out eval pass rate, stock analysis
- +0.29
- ranking's correlation with abnormal volume
- Python
- Claude API
- Flask
The Language of Flowers
A private tarot site, my little tarot, with a 78-card floral deck I generated and composed, readings by Claude in Chinese or English, and an eval that holds those readings to the deck and to the way the question was asked.
- 78
- cards in one style, chosen from 577 generated images
- 98%
- readings passing every check, Opus 5.5 (60 graded)
- 2
- languages, switchable mid-reading without reshuffling
- Node.js
- Claude API
- Gemini image generation
Research
Mar – May 2026
Entropy-Guided Generative Verifier for LLM Reasoning
Token-level uncertainty and generative verifiers for picking reliable reasoning paths on GSM8K and MATH500.
- LLM reasoning
- Uncertainty
- Best-of-N
Jun – Aug 2024
Feature Contribution in KNN Models with Shapley Values
Data valuation with Shapley values: removing points around ±2σ improved KNN accuracy.
- Shapley values
- Data valuation
- KNN