Featured Project Case Studies

Challenge

Create a repeatable multi-source shot-data pipeline that supports stable downstream modeling and archetype analysis without schema drift.

Process

• Ingested and merged 1.1M+ shots across five seasons from multiple sources
• Applied schema contracts and validation before feature generation
• Implemented rerunnable SQLite load/upsert behavior for repeatable refreshes
• Engineered xFG, residual, and SDI features for GAM and clustering workflows

Outcomes

• xFG model reached 62.7% holdout accuracy
• Generated stable feature layers for shot-quality and value analysis
• Delivered interpretable archetype visuals and residual diagnostics

Challenge

Unify chat, PDF Q&A, and real-time external tools into one secure workflow with consistent authentication and data isolation.

Process

• Split system into React/Vite frontend and Django REST backend
• Implemented JWT access and refresh token lifecycle
• Integrated Gemini via LlamaIndex ReAct with tool orchestration
• Built PDF index persistence and lazy query loading

Outcomes

• Single user workflow for chat, docs, and tools
• Composable API surface for fast feature iteration
• Reliable user-scoped data separation across notes and history

Challenge

Build a robust R package that can ingest and validate 1,000+ games per season with graceful failure handling.

Process

• Implemented parallel collection wrappers with retry logic
• Created manifest checks for missing, postponed, and canceled games
• Enforced schema maps and type-casting contracts
• Built idempotent upsert tables and migration hooks

Outcomes

• Produced four stable parsed core tables
• Established 13 fixture-based tests without live API dependency
• Enabled repeatable incremental updates without duplicate rows

Challenge

Audit AI and ethics course coverage where catalog data is distributed across term pages, prefix lists, and inconsistent page structures.

Process

• Extracted prefixes from PDF and maintained reusable prefix lists
• Built Selenium crawl flow with resumability and empty-prefix logging
• Separated analysis into precision, broad recall, and ethics scripts
• Exported cleaned CSV outputs and report-ready summaries

Outcomes

• Scraped 12,944 records across 150+ prefixes and two terms
• Created transparent precision-vs-recall reporting in R Markdown
• Delivered inspectable stage-by-stage transformation outputs

Challenge

Model home-team win probability across 23 seasons while preventing temporal leakage that would inflate metrics.

Process

• Built lagged rolling features across 3/5/10-game windows
• Engineered possession metrics and matchup differential features
• Compared ridge, elastic net, logistic, and random forest models
• Integrated Elo variants and calibrated holdout thresholds

Outcomes

• Best ridge model achieved 67% accuracy and 0.684 AUC
• Maintained chronological train/test integrity
• Documented reproducible model comparison workflow

Challenge

Deliver four progressive deep-learning projects while preserving reproducibility and making results auditable.

Process

• Implemented ANN baseline workflows and regression tests
• Built HPO model-factory patterns with W&B experiment tracking
• Developed CNN interpretation through filter visualization
• Implemented RNN sequence pipelines with custom tokenizer

Outcomes

• Completed ANN, HPO, CNN, and RNN progression
• Established reproducible training utilities and shared seeds
• Produced tracked experiment outputs for comparison and reporting

Challenge

Ship a production-ready freemium analytics platform with secure auth, billing, and AI-style Q&A without exposing raw SQL or sensitive internals.

Process

• Built Fastify API modules for auth, routes, billing, and Q&A
• Implemented regex-intent Q&A templates to avoid user SQL execution
• Added Stripe checkout and webhook subscription management
• Hardened with JWT policy, Zod validation, rate limits, and Redis caching

Outcomes

• Reached 69 passing tests across 25 files
• Achieved A+ security posture in project audits
• Delivered reliable freemium controls and optimized dashboard UX

Challenge

Design a multi-sport expected points framework that explains shot quality and value while remaining interpretable to both technical and non-technical audiences.

Process

• Built xFG and xG workflows with POE and residual diagnostics
• Ran GAM tensor-surface models with PDP interpretation
• Clustered player shot profiles via GMM and PCA
• Combined salary collection with POE per million value rankings

Outcomes

• Produced interpretable cross-sport shot-quality metrics
• Built reusable scripts for calibration, clustering, and value analysis
• Delivered an interactive Streamlit capstone dashboard

Challenge

Build a Dockerized ingestion system that can rerun safely at scale while preserving auditability and failure tracing.

Process

• Parallelized HTTP fetches with concurrent.futures
• Mapped variable ESPN JSON schemas into relational structures
• Implemented chunked ON CONFLICT upserts via SQLAlchemy
• Automated operations through Makefile commands and Docker compose

Outcomes

• Established idempotent PostgreSQL ingestion behavior
• Created failure-stage observability through ingest-failure tracking
• Enabled repeatable local and containerized execution workflows

Challenge

Translate highly technical malware behavior into a rigorous, evidence-based analysis that remains readable and academically structured.

Process

• Mapped propagation and exploit chain details across four zero-days
• Analyzed certificate abuse and PLC hardware-targeting logic
• Documented attack-vector mechanics and sensor spoofing behavior
• Structured findings in publication-style R Markdown and LaTeX

Outcomes

• Delivered a comprehensive Stuxnet technical report
• Connected exploit mechanics to real-world infrastructure impact
• Showcased cybersecurity depth aligned with minor focus