Glenn Dalbey - Resume

Data Scientist with an MS in Data Science who builds things that actually work. I enjoy building and training models on my own multi-GPU homelab, deploying new techniques, and discovering unique ways to solve problems. Whether it's analysis, research, or projections, I dig into how something ticks, figure out why, and extract information that drives better outcomes. My multi-node homelab stack lets me run production applications, train models, and run MLflow, all on my own hardware. 20+ projects on GitHub covering deep learning, medical imaging, NLP, computer vision, and analytics.

NFL Big Data Bowl 2026 - Kaggle Bronze Medal

• Bronze Medal in prestigious Kaggle competition predicting NFL player trajectories from tracking data
• Conducted systematic exploration of 847+ experiments across 15+ neural network architectures
• Best single model: 6-Layer Spatial-Temporal Transformer achieving 0.547 Public LB score
• Best ensemble: 3-model blend (ST Transformer + CNN + GRU) achieving 0.540 Public LB with architecture diversity
• Engineered 167 features including kinematics, ball-relative positions, temporal patterns, and geometric features with Voronoi tessellation
• Implemented novel geometric attention with spatial distance modulation and frozen encoder fine-tuning
• Utilized multi-GPU training, mixed precision (FP16), and test-time augmentation for +0.005-0.010 improvement

• Trained 105 deep learning models (21 architectures × 5 folds) for CT angiography aneurysm detection
• Tested 51 ensemble configurations; best ensemble META_E_top3_weighted achieved AUC 0.8624
• Key discovery: Smaller models (SE-ResNet18) statistically outperform larger models on limited medical data (r=-0.42, p<0.01)
• Built complete pipeline: DICOM→NIfTI→ROI extraction→Training→Ensemble across 4 GPUs simultaneously
• Multi-label classification across 14 classes with severe class imbalance handling (1.2% to 42.8%)

• Built and deployed live production healthcare AI triage system with sub-second response times
• Achieved 93.8% combined multi-modal accuracy and 98.0% burn classification accuracy
• Implemented 5-model ensemble architecture combining DistilBERT (NLP) and CNNs (Computer Vision)
• Successfully handled extreme class imbalance (29.7:1 ratio) with advanced loss functions
• Built production pipeline with Flask API, AWS S3 integration, and comprehensive safety protocols

• 7 statistical methods + 3 ML models applied to 41,200 cases across 101 years (9,204 county-decade combinations)
• Kenedy County, TX: 46.86σ composite z-score persists after Bayes shrinkage, OLS, Random Forest, and FDR correction
• I-35 corridor: 170% increase in missing persons (193 to 521 cases), structural break at 2020 (p < 0.001)
• Spatial autocorrelation confirmed (Moran's I = 0.22, z = 26.03) with LISA hotspot clustering along TX border
• Validated against known serial killers (Ridgway: 4.38σ, Gacy: 1.34σ)
• Live 7-page Streamlit dashboard with geospatial visualization and ARIMA 5-year forecasting

• Agentic 3-call LLM pipeline that plans, executes, and synthesizes 22 statistical methods across Census and CMS data
• First pass planned 9 analyses, follow-up requested 5 more for 14 total; full pipeline cost about 6 cents in LLM tokens
• Market scored 48.3/100 overall (demand pressure 53.6, competitive position 40.3) across 10 nursing facilities with 896 beds
• Chi-square confirmed significant link between ownership type and quality rating (p = 0.0067)
• 622-word executive briefing with source citations on every claim, traceable to Census Bureau and CMS Care Compare

• Developed predictive model achieving 90.9% ROC AUC on future data validation for 1000+ yard seasons
• Reduced overfitting gap from 18.5% to 0.4% (97.8% reduction) through feature optimization (46→20 features)
• Implemented temporal validation strategy ensuring model generalization to future NFL seasons
• Created production-ready ensemble model for NFL draft analysis with comprehensive data pipeline (2006-2024)

• Originally prototyped in Python, rewritten in C++17 for performance — evolved through 4 major versions with systematic optimization
• Indexes 2.45 billion tokens across 5M+ chunks with 0-42ms search (avg 12ms, down from 500ms in v1) — CPU-only, minimal RAM
• Dual LZ4/Zstd compression, 15 REST API endpoints, 47 tests at near 100% accuracy; conversations and files continuously indexed
• Building LLM chat interface and MCP tool for terminal-based AI workflows and project knowledge management

• Three-tier LLM routing (local Ollama to cheap API to premium models) reducing costs by routing ~80% of requests to free local models
• Triple-layer RAG memory: BM25 over 5M+ chunks, ChromaDB semantic search, Redis session cache with graceful degradation
• 18 MCP-compatible tools and multi-agent sandbox where 4 LLM personas deliberate at zero API cost
• 4,200+ lines of async Python on FastAPI for a private multi-node GPU cluster

• Designed and operate multi-tier AI homelab: dual RTX 5090 training node + RTX 3090 Ti/3090 secondary node
• Built 256GB unified memory LLM inference cluster (2× Ryzen AI Max+ 395) running Kimi K2, Qwen 3, GLM 4.6
• Implemented automated active-memory pipeline with n8n orchestration, RAG storage, and hot/warm/cold tiering
• Deployed Proxmox VE backbone with pfSense firewall, VLAN segmentation, and 10Gb networking (100TB+ storage)

• Provide data science and analytics consulting services for business intelligence initiatives
• Develop AI-powered Excel analytics platform enabling natural language data queries
• Design custom analytical solutions and machine learning models for client-specific challenges
• Support data-driven decision making through advanced analytics and predictive modeling

• CI Department Representative leading process improvement and operational efficiency initiatives
• Developed comprehensive training curriculum for warehouse personnel, improving onboarding efficiency
• Designed and implemented the Zones Project, modernizing material flow training systems
• Led departmental CI mapping initiatives to improve operational efficiency and reduce cycle times
• Optimized material replenishment processes using bin methodology, reducing operational inefficiencies
• Managed supply chain logistics and SAP-integrated inventory management
• Supported engineering teams in workflow re-splits and cycle time analysis for production optimization