Harry (Jiajun) Hu

• Built and deployed a geospatial ML outlier-detection pipeline (Isolation Forest, DBSCAN + tree-based models) to rank rooms for expert review — uncovered 14% under-utilized space and projected 9% property revenue uplift.
• Integrated Zillow API to collect real-time valuations for 220 properties and ran geospatial analysis in ArcGIS Pro — revealing over 10% latent revenue upside to guide long-term space-use strategy.
• Optimized form UX via A/B testing, cutting auditor input time by 25% and halving error rates.
• Built live Power BI dashboards and delivered weekly insight reviews, raising daily audit coverage by 20%.

• Fine-tuned YOLO with frozen CNN backbone on limited blueprint data, chained OCR + rule-based aggregation to compute linear meters — reducing takeoff from 60 min to 30 sec at 97% accuracy.
• Designed domain-specific augmentation (rotation, scaling, line-width/contrast perturbation) that boosted minority-class recall by 20% under severe class imbalance.
• Identified YOLO's architectural ceiling on dense small targets with structural relationships — reflection that directly informed the segmentation pivot in my next project.

• Built an end-to-end ML pipeline combining OCR, LLM-assisted semantic extraction, and data-centric optimization to classify 50,000+ historical documents — achieving 87% precision and 93% recall on 3,000 labeled PDFs.
• Designed a lightweight labeling QA loop and data quality controls that improved training signal without large-scale data collection — a data-centric approach over model complexity.
• Implemented LLM-assisted outlier detection to flag edge cases and rare document types, preventing noisy samples from degrading model performance.
• Packaged a repeatable inference workflow and delivered stakeholder reports translating model predictions into actionable research insights.

• Decomposed multi-stage ad funnel (impression → click → conversion → revenue) to identify bottleneck segments, driving budget reallocation that lifted ROAS by 18%.
• Designed and supported A/B tests on creative layouts and ad placements with proper sample sizing and significance thresholds.
• Developed time-series analyses over rolling windows (7/15/30 days) to identify seasonality patterns and optimize campaign timing.
• Built Power BI dashboards with ETL pipelines for self-serve insights, enabling non-technical stakeholders to act without analyst support.

Washington University in St. Louis, McKelvey School of Engineering

New York University, College of Arts & Science