Wesley Tao

PayPal

Staff ML Scientist (Data Scientist → MTS1 MLE → Manager → Staff ML Scientist)

San Jose, CA

Customer-Facing PayPal Assistant Agent | SLM + Tool Calling

• Integrated an in-house open-source SLM challenger into a production customer-service digital agent (disputes, payment holds/declines, account limitations) with tool-grounded execution; challenger in employee A/B testing with comparable quality at materially lower p90 latency and cost.
• Single-threaded ML owner for the SLM track; drove key architecture tradeoffs (single vs. multi-agent, tool adapter strategy, MCP vs. direct APIs, transcript-based evaluation) with Eng/PM and translated decisions into a shippable plan.
• Shaped the multi-agent "swarm" approach: central orchestrator/router delegates to domain sub-agents; scaled to 8 sub-agents (10–12 tools each) with a roadmap to 30+ sub-agents and 108 legacy API conversions.
• Owned Manage Bank sub-agent end-to-end (tool schemas, synthetic/fine-tuning data, offline simulation, prompt/model iteration), exceeding production baseline task success.
• Built offline evaluation + regression stack via transcript replay/user simulation; defined pre-prod gates for task success, tool-call correctness, hallucination control, and grounded synthesis.
• +12% task success across 8 agents via rule-driven synthetic data + targeted fine-tuning; reduced p95 single-turn latency 6.0s → 1.2s via model compression and consolidating multi-call flows into single-call execution.

Teammate Smart Reply (Agent Assist for Customer Support)

• Led GenAI Smart Reply from prototype to production, reducing average handle time (AHT) by ~10s and recontact rate by ~10% through iterative model/prompt and product tuning.
• Improved Smart Reply coverage by +7% by expanding RAG sources to include customer-facing email/chat templates; used LLM summarization to normalize content into agent-style replies.
• Built retrieval evaluation for RAG using gold human-labeled query→doc relevance and silver click-through labels; tracked nDCG and offline regressions to tune retrieval quality.
• Built an end-to-end feedback and evaluation pipeline (adoption, human edit distance, regression test pass rate), improving iteration safety across prompt/model releases.

Patents & Other Projects

• Patent filed (Dec 2025): System and Method for Conversational Workflow via an AI-Driven Schema — schema-driven, multi-turn conversational workflows for reliable tool-grounded task completion.
• Fine-tuned GPT-3.5 for PayPal Assistant using RAFT-style data strategies (2023); achieved near-GPT-4 quality on hallucination rate and reply quality, delivering $3M+ annualized savings; patent filed.
• Explainable AI (XAI) for PayPal's flagship Seller Fraud Models (2022): built tree-SHAP and two-layer reason code mapping; adopted in production enabling 678K txns/yr, 31K contacts savings, and $95K margin uplift.

Ushur Inc

Data Scientist

Santa Clara, CA

Developed an NLP (natural language processing) pipeline for insurance underwriter’s decision automation process. Integrated a dockerized rule-based expert system UMLS (unified medical language system) for feature extraction which greatly improved coverage for disease detection from 40% to 80% Implemented a tfidf and SVM (support vector machine) model for email classification and visualized the confidence scores distribution for unseen categories which beats the production model in terms of overall accuracy and robustness. Designed a statistical brand proximity metric which evaluate the product’s user engagement and efficiency of the system response; A nonprovisional patent being applied in progress

Pactera Oneconnect AI Lab

Machine Learning Researcher Intern

New York, US

Built an end-to-end chatbot assistant to facilitate the company’s hiring process collaborating with other engineers Implemented LSTM, SVM, Tree-based models to upgrade hard-coded dialogue and created a user simulator to generate simulation data for testing Created and maintained the SQL databases on AWS for chatbot to access and retrieve relevant information

Adatos.ai

Data Scientist Intern

Implemented a Deep Learning, powered predicting model of palm tree yield (tree detection, tree counting/density estimation, leaf and soil nutrient analysis, fertilization analysis, age estimation and weather/disasters analysis) on satellite imagery to incorporate the signals in palm oil commodity future trading strategy

Institute of Data Science, Fudan University

Research Assistant

Completed independently a research project on electricity user's behavior study with Hausman-Taylor model and test effectiveness of the electricity pricing policy in Shanghai.

Tested and proved the hypothesis that even at a low price difference +/- $0.03/kWh, people under the non-flat rate policy tend to use 12% less after 22:00 (peak hour)

Used Hausman-Taylor model to exclude the unobserved individual effect and successfully measured price elasticities