职业经历 - Harvey's Personal Website

Full Time Job

2025-12 → Now

Senior Data Scientist

IntuitDS

Real-time Next Best Page Recommendation SystemBuilt a system for QuickBooks.com to surface the highest-converting landing page for each prospect. Designed a two-stage architecture pairing an LLM-based relevance-analysis agent with a causal uplift model for conversion scoring, taking their intersection as the final recommendation. A/B testing showed positive lift across signup and GNS, driving an estimated ~$20M annual incremental LTV, with the model now deployed across 100% of US desktop traffic.
SINGULAR Knowledge Layer (Patent Filed)Co-architected SINGULAR, an internal MCP server that unifies data knowledge across every domain into one governed layer, paired with Intuit's Databricks MCP so agents go straight from knowing what to query to executing it. Filed a patent application for SINGULAR, now broadly adopted by DS, marketing, and product teams org-wide to run complex analysis within Claude, Cursor.

Python
SQL
LLM Prompt Engineering
Causal Inference
Uplift Modeling
A/B Testing
Predictive Modeling
MCP (Model Context Protocol)
Semantic Layer Design
TypeScript
BM25 Search
Databricks

2023-06 → 2025-12

Senior Data Scientist

IBMDS

Arizona State DCS Virtual AgentBuilt a policy-focused virtual agent for Arizona State DCS caseworkers to deliver instant, accurate answers to policy questions. Implemented a Retrieval-Augmented Generation (RAG) architecture on Azure AI Search and Azure App Service to ground responses in the agency's policy corpus, delivered through a web front end integrated into caseworkers' Microsoft 365 workflow. The tool reached human-expert-level accuracy post-launch, significantly improving case resolution efficiency.
AI Agent Content GeneratorDeveloped an AI agent-based content generation system to automate pharmaceutical regulatory documentation and compliance reporting. Orchestrated a multi-agent pipeline (creator, scorer, reviser) with the AutoGen framework, exposed as containerized FastAPI microservices and deployed on Azure Kubernetes (AKS). The system generated 95%-accurate compliance reports and marketing pages while reducing manual processing time by 75%.
Nashville PD Video AnalyticsDeveloped a generative AI solution for the Nashville Police Department to automate analysis of body camera video footage. Built an AWS Bedrock-powered multimodal pipeline for face recognition, transcription, and LLM-based summarization and entity extraction, integrated with IBM Appian for case-management workflows. The solution converted raw footage into structured, searchable intelligence, saving officers hundreds of manual review hours monthly.
Hyatt Sales AutomationDeveloped an AI automation system for Hyatt's sales team to generate RFP responses and sales proposals automatically. Orchestrated AWS Claude 3 via LangChain for structured proposal generation, rendering dynamic PDFs from templated layouts and deploying as containerized services on AWS ECS. RFP response rates rose from 5% to 100%, saving sales managers 80+ hours of manual work monthly.

PyTorch
TensorFlow
Scikit-learn
Transformers
LLM Fine-tuning
Prompt Engineering
RAG
Agentic AI
LangChain
LangGraph
Encoder training
AWS
Azure
Docker
Kubernetes
HTML
CSS
JavaScript
Django
SQL
Streamlit
Agile

2020-10 → 2022-04

Data Scientist

Kuai ShouDS

Machine Learning Prediction AnalysisLed onboarding strategy for creators transitioning into livestreaming to grow platform revenue. Built predictive models to forecast creator earnings and design tiered incentive strategies segmented by predicted value. A/B testing demonstrated a $5M monthly revenue increase over the prior incentive structure.
Causal Influence AnalysisPartnered with the operations team to reduce creator churn (attrition) on the platform. Applied feature engineering and SHAP analysis to identify key churn drivers, including early-stage follower growth, then designed and validated retention strategies through A/B testing. The strategies reduced long-term churn by 20% and were adopted as standard practice.
Heterogeneity AnalysisCollaborated with top influencers (KOLs) to improve retention and spending among young users on the platform. Applied propensity matching and difference-in-differences (DID) analysis to isolate the causal effect of KOL content on engagement, then used uplift modeling to quantify heterogeneous effects across user segments, revealing the strongest impact among high-value male users. These findings, aligned with the company's target demographic, supported a $10M increase in KOL partnership investment.
Experiment AnalysisOwned experiment design to evaluate new business strategies for new users and creators. Designed and monitored A/B tests in real time, applying quasi-experimental methods, including synthetic control and matching, in settings where randomization was infeasible. Synthesized results into go/no-go recommendations that shaped the rollout of new growth initiatives.