Hi, I'm Yifan I turn dataideascuriositychaos into stories.

Taking Initiative & Solving Bottlenecks

During my internship at ByteDance, I noticed a real problem: the Strategy Team was overwhelmed by the need to manually read and tag thousands of macroeconomic reports to track market sentiment. It was slow, subjective, and inefficient.

Instead of waiting for someone else to fix it, I designed and deployed an automated scoring system for the "Macro Daily" Feishu Push Assistant using Large Language Models.

What I Built

• Created a testing framework to refine prompt configurations through multi-version comparison, ensuring scoring matched professional human judgment
• Processed and analyzed over 3,000+ macroeconomic articles
• Generated automated monthly sentiment trend charts and thematic heatmaps the team now uses for decision-making
• Built visual dashboards using the internal BI tool "Fengshen" for the Macroeconomic Monthly Report
• Conducted focused analysis on trade, total social financing (TSF), and money supply (M1, M2)

Why This Matters

This wasn't a class assignment. It was a real business tool used by a real strategy team. It showed me that identifying workflow inefficiencies and designing scalable solutions is just as important as technical skills.

Tech Stack

The Real Challenge

In most data science courses, datasets come pre-cleaned and simplified. I wanted to challenge myself with the messy reality of federal data to answer a complex question: Does air quality genuinely impact housing prices?

I consolidated sixteen separate years of raw CSV files from the EPA, FHFA, and BEA, handling non-standard missing values and sanitizing inconsistent formatting that would have broken a standard model.

Key Findings

• Cleaner Air, Higher Prices: Better air quality (lower AQI) is associated with significantly higher housing prices
• Independent Effect: Controlling for income and other factors, the effect remains statistically significant (coefficient: 0.389): air quality affects housing independently of wealth
• Policy Implication: Environmental improvements generate direct asset value appreciation, providing economic justification for stricter regulations

Why This Matters to Me

This project represents my capability to manage the entire data science lifecycle, not just the final modeling stage. It demonstrates the technical resilience to wrangle unstructured information into clean, analytical format, and translate chaotic real-world data into clear, policy-relevant economic insights.

Project Walkthrough

Tech Stack

Born from Personal Frustration

As a non-native speaker preparing for the GRE, I found that existing tools to learn vocabularies were rigid. Traditional PDFs lacked interactivity, and mainstream apps didn't allow me to upload my own custom vocabulary lists.

Instead of waiting for a better tool, I decided to build one: a web app designed to help non-native speakers master GRE synonyms. I adopted a Product Manager mindset, identified a specific gap in the market for international students, and leveraged AI-assisted coding to rapidly prototype and deploy it.

Key Features

• Smart Dashboard: Progress tracking with daily and cumulative word counts, plus an auto-resetting daily counter
• AI-Powered Story Generator: Creates short narratives using your recently studied words with customizable themes
• Smart Excel Import: Upload custom .xlsx vocabulary files, with no dependence on pre-built lists
• Immersive Flashcards: 3D flip card interface with star-marking for difficult words
• Privacy First: A bring-your-own-key system where API keys stay in your browser and are never transmitted

App Screenshots

Why This Matters

This project is proof that I can turn a personal observation into a functional solution. It demonstrates market insight, ability to execute, and the willingness to look beyond lines of code to understand user pain points.

Tech Stack

The Problem

Accurate precipitation prediction in the dry California Central Valley is vital for regional agricultural production. Farmers need to know when extreme rainfall will hit, not just seasonal averages, but the actual storms that matter.

What I Did

• Extracted 10-year weekly precipitation data (2016-2025) revealing a highly right-skewed distribution (Skewness = 2.85)
• Applied log1p transformation to stabilize variance
• Identified a clear 52-week annual cycle for seasonal modeling

Models Compared

• LSTM (Long Short-Term Memory): MAE 0.9675, RMSE 1.7596: best at capturing extreme precipitation peaks
• BSTS (Bayesian Structural Time Series): MAE 0.72, RMSE 1.45: accurate for weekly trends but too "smooth" for extremes
• SARIMA: MAE 2.92, RMSE 4.76
• ARIMA: MAE 3.04, RMSE 4.30

Key Finding

While BSTS is accurate for seasonal trends, it is too conservative for risk management. LSTM is the superior model because it captures the stochastic volatility and precipitation peaks that other models miss. Since predicting storms is more critical for agriculture than tracking averages, LSTM is the right architecture for high-stakes forecasting.

Validation

5-year Rolling Cross-Validation (2021-2025) with Log-Min-Max preprocessing and Adam/MCMC optimization.

Research Poster

Click to view full size

Tech Stack

Overview

At Bayer I built three AI-assisted workflow prototypes for medical and regulatory documents. Each one takes study information that arrives as loose, unstructured text and turns it into a structured draft that an expert then reviews and signs off on.

Three Separate Tools

These are three different tools for three different jobs. They do not share a pipeline, and each one needed its own logic and its own checks.

Tech & Focus

The Question

Analyzed an NHANES-derived dataset of 7,719 adults to investigate the association between self-reported hypertension and weak or failing kidney function, with age and sex considered as potential confounders.

Method

• Applied t-tests, chi-square tests, and multivariable logistic regression
• Adjusted for age and sex to isolate the association
• Critically evaluated confounding, self-report bias, and the limits of causal inference in a cross-sectional design

Key Finding

After adjusting for age and sex, participants with hypertension had 3.17× the odds of weak or failing kidneys (95% CI: 2.30 - 4.41) — a strong, statistically significant association.

Research Poster

Click the poster to open the full-size PDF

Tech & Focus

What I Led

• Led a 4-member research team studying the mid-term investment potential of cloud-computing industry funds
• Built a time-series database of 6-month fund NAV, volume, and volatility
• Applied regression models (OLS, Ridge, and Lasso) and correlation analysis
• Handled data cleaning and visualization in Python (pandas, statsmodels, matplotlib, seaborn)

Final Presentation

‹ ›

1 / 25

Use the arrows to flip through all 25 slides, or open the full deck as a PDF.

Tech & Focus