Yiming (Vivian) Jia
858-***-**** ******@***.*** LinkedIn Github
EDUCATION
New York University, Center for Data Science Sep 2024 - May 2026 M.S. in Data Science GPA: 3.79/4.00
University of California, San Diego Sep 2020 - Jun 2024 B.S. in Applied Mathematics, Minor in Data Science GPA: 3.83/4.00 SKILLS
Languages: Expert – Python, SQL (PostgreSQL, MySQL, Oracle); Intermediate – Java, JavaScript AI & Machine Learning: PyTorch, Transformers, scikit-learn, XGBoost, LangChain, LangGraph, Pydantic, RAG, AI Agents, Tool Calling, Prompt Engineering, Structured Outputs, LLM Evaluation, Vector Databases, Guardrails, Human-in-the-Loop, Reinforcement Learning, Contextual Bandits, Feature Engineering, Model Evaluation, Cross-Validation, Imbalance Handling Data & Cloud: Pandas, NumPy, PySpark, BigQuery, Vertex AI, AWS, REST APIs, ETL/ELT Developer Tools & Infrastructure: Git, GitHub, Linux, Shell, Monitoring & Observability Analytics: A/B Testing, Experiment Design, Hypothesis Testing, Tableau, Power BI WORK EXPERIENCE
eWorld Enterprise Solutions, Inc. New Jersey, USA
Applied AI Data Scientist Sep 2025 - Dec 2025
l Designed and implemented an AI-assisted Hawaii SNAP evidence-verification workflow with Python and Gemini 2.0 Flash on Vertex AI, converting PDF applications and interview notes into member-level JSON checklists; achieved 96% case-level accuracy across 200 internally reviewed cases and reduced manual follow-up effort by 87%. l Integrated Gemini 2.0 Flash with Python through Vertex AI APIs, building reusable components for LLM requests, response parsing, and downstream orchestration to support automated evidence-verification workflows. l Configured category-specific policy queries, prompts, and structured-output schemas in a RAG workflow, grounding checklist decisions in approved policy passages; implemented rule-based validation, contradiction checks, policy-reference validation, and human-review fallback, reducing unsupported or invalid outputs from 24% to under 3%. l Built reusable Python preprocessing components to extract fixed PDF fields, normalize household-member data, and combine application data with source-labeled interview notes into standardized case inputs, enabling consistent downstream identification of missing or conflicting evidence. Genmab New Jersey, USA
Data Scientist Sep 2025 - May 2026
l Trained and compared Logistic Regression, Random Forest, Gradient Boosting, and XGBoost models to predict mortality risk across 1,032 CRS reports, addressing class imbalance and achieving ROC AUC 0.72 / PR AUC 0.59 to prioritize high-risk safety reports.
l Built a configurable pipeline integrating 6K+ PubMed publications with FAERS data for safety-signal validation; identified a 98% increase in CRS reporting while publication volume remained stable. Antalpha New York, USA
Data Scientist Jun 2025 - Aug 2025
l Developed a Python telemetry scoring workflow using Logistic Regression and lagged hashrate, temperature, and power features to identify sustained ASIC miner underperformance, achieving ROC AUC 0.81 and reducing non-actionable performance flags by 21% compared with threshold-based rules in historical evaluation. l Automated miner telemetry ingestion and preprocessing through the Foreman API and built a reusable monitoring dataset for Tableau dashboards tracking hashrate and operating efficiency, reducing weekly data preparation time by over 90%. Qifu Technology Inc. Beijing, China
Data Scientist Jul 2023 - Sep 2023
l Developed an XGBoost 7-day borrowing propensity model using lending, repayment, and behavioral features, achieving ROC AUC 0.82 and supporting campaigns that increased loan issuance by 12% and reduced acquisition cost by 10%. l Built a Python-orchestrated SQL automation workflow that executed daily analytical queries, generated refreshed reporting tables, and replaced manual recurring calculations, reducing reporting runtime from 2 hours to 20 minutes. PROJECTS
LLM Reasoning Strategy Selection with Contextual Bandit Optimization Mar 2026 - May 2026 l Designed a Python-based LLM evaluation pipeline to generate, cache, validate, and benchmark three prompting strategies across 500+ reasoning problems, enabling reproducible comparison of accuracy, reliability, and inference cost. l Formulated prompt strategy selection as a contextual bandit problem that balanced reasoning accuracy against inference cost, reducing reasoning cost by 29% while remaining within 1.6 percentage points of the strongest fixed strategy. l Implemented automated output validation with final-answer extraction, mathematical-expression normalization, SymPy verification, and parsing diagnostics, reducing invalid outputs by 38% and improving evaluation reliability. Language-Guided Exploration in Reinforcement Learning Apr 2026 - May 2026 l Investigated whether language-like state abstractions improve reinforcement learning exploration and zero-shot transfer by developing a PyTorch PPO training pipeline for MiniGrid procedural environments. l Conducted 90 controlled experiments across 4 environments and 6 exploration settings, demonstrating that medium- and coarse-grained state abstractions improved held-out transfer performance in 3 of 4 environments.