John He
AI Engineer & Data Scientist
Specializing in LLM Agents, RAG Systems, and Scalable ETL Pipelines.
MS in Machine Learning and Data Science at Northwestern University
Specializing in LLM Agents, RAG Systems, and Scalable ETL Pipelines.
MS in Machine Learning and Data Science at Northwestern University
Overview: Built an agentic RAG system over 800,000+ arXiv abstracts with query classification and semantic search.
Key Results: Query classifier routes to direct answers or a RAG-enhanced flow with a query rewriter, document retrieval, and answer generation. Returns relevant abstracts and a fine-tuned 1.8B-parameter LLM abstract simplifier to explain technical jargon. Deployed as a Streamlit chatbot.
Overview: Designed cloud-native data infrastructure processing 2.5M+ network traffic records for intrusion detection.
Key Results: Implemented Hive-style date partitioning in S3 and used AWS Glue Crawlers for automated schema discovery. PySpark ETL jobs in Glue transformed raw PCAP CSVs into a Redshift-ready schema; a Random Forest classifier was trained as part of the pipeline and served via a containerized FastAPI endpoint on ECS Fargate.
Overview: A RAG system for question answering tasks deployed on Railway using Pinecone as the vector database and Grok for natural language processing.
Overview: An automated data pipeline for fetching and processing financial data from Yahoo Finance using Fivetran to extract data from Yahoo Finance, Snowflake for loading/storage, dbt to transform it, and Streamlit to visualize it.
Key Results: Led 3 person team and delivered 3 analytics tools (lineup predictor, synergy finder, tactical chatbot) in a 48-hour hackathon.
Overview: Comparative analysis of five ML models predicting customer churn on 165K records, with strategic threshold tuning to optimize retention interventions.
Key Results: LightGBM achieved 86.54% accuracy; threshold tuning boosted at-risk customer identification from 55% to 75% recall.
Key Results: Forecasted sales for top 10 departments, identifying those with >90% projected ROI to guide strategic investment.
Automating user queries via multi-tool calling and RAG.
Established significant health indicators using Cox PH models and survival analysis.
Reduced manual review time by 90% (30 min → 3 min) using LangChain & HPC.
Cut experimental sample size by 33% via stratified sampling.
Overhauled research data infrastructure, doubling accessible data volume.
Northwestern University
Middlebury College