Analytics & technical infrastructure portfolio.

Business questions

• How are A/R loads and collection rates tracking month over month?
• Which payers generate the highest write-off burden per encounter?
• Where are denial rates concentrated, by provider, CPT, or payer?
• How does reimbursement compare to allowed revenue across encounter types?

Approach

• Cleaned and transformed raw billing data using Python and Power Query
• MPRR reductions and standardized adjustment coding
• Built univariate and multivariate analysis across payer and encounter dimensions
• Designed an interactive Excel dashboard with slicers for payer, month, and provider
• Produced a comprehensive billing report with provider-level KPIs

Key metrics

• Reimbursement totals and reimbursement rate
• Denial volume and denial trends
• Operational efficiency indicators
• Profitability comparison

Outcome

Produced a reporting tool that enables comparison of performance drivers and demonstrates how reimbursement patterns, denials, and operational processes influence financial outcomes.

Objective

Develop a computer vision model capable of accurately classifying facial expressions into four emotion categories: happy, sad, surprised, and neutral. Part of a broader direction in affective computing and emotionally intelligent machine behavior.

Approach

• Baseline ANN model established as performance reference
• Progressively refined CNN architectures across four model iterations
• Applied dropout regularization to reduce overfitting
• Increased training epochs to improve feature refinement and generalization
• Evaluated grayscale vs. RGB input strategy for transfer learning alignment
• Libraries: TensorFlow, Keras, NumPy, pandas, Matplotlib, Seaborn, scikit-learn

Key metrics

• Baseline ANN: 56% F1 score (best on surprise)
• Model 1 (simple CNN, no dropout): 59% F1 score
• Model 2 (refined CNN): 69% F1 score
• Model 3 (optimized CNN): 91% F1 score
• Evaluated on happy, sad, neutral, and surprised categories

Outcome

Achieved 91% classification accuracy on facial emotion recognition through iterative model improvement. The project demonstrated that CNN depth, dropout tuning, and epoch scaling significantly outperform simpler ANN architectures for image-based emotion classification. Key insight: matching input format (grayscale) to model expectations improved alignment and reduced preprocessing artifacts.

Objective

Design and operate a centralized, production-grade self-hosted platform capable of running multiple public-facing and private services under a unified domain, with enterprise-level routing, security, and access control.

Architecture

• Docker-based service isolation and container orchestration
• Nginx reverse proxy with subdomain-based routing across all services
• Cloudflare for DNS management, DDoS protection, and traffic proxying
• Let's Encrypt for automated SSL certificate provisioning and renewal
• WireGuard VPN for encrypted remote access and secure tunneling
• Pi-hole for network-level DNS filtering and ad blocking
• MariaDB for persistent relational data storage across services

Services

• huy.gg: production website and public portfolio
• Self-hosted email infrastructure
• ownCloud for private file storage and sync
• Jellyfin for media streaming
• Additional containerized services under active management

Outcome

Delivered a fully operational, multi-service production environment that runs continuously under a live public domain. Demonstrates end-to-end systems thinking across networking, security, containerization, and infrastructure management, well beyond the scope of a typical homelab.