Derek Lankeaux

Data Scientist and Applied Statistics M.S. candidate focused on experimental design, Bayesian inference, machine learning, and LLM evaluation.

LinkedIn · GitHub · Portfolio · Résumé

Portfolio

This repository contains four independent technical case studies. Each project links to a full Markdown report and a publication-formatted PDF; the summary below is intended to make the evidence and the scope easy to evaluate.

Project	Question	Methods	Evidence
AI Safety Red-Team Evaluation	How can safety evaluation scale beyond manual review?	LLM ensemble annotation, supervised classification, Bayesian risk analysis	Report · PDF · Project page
Breast Cancer Classification	Which ensemble methods perform well on WDBC diagnostic features?	Benchmarking, calibration, feature selection, explainability	Report · PDF · Project page
LLM Ensemble Bias Detection	Can multiple LLM judges support uncertainty-aware content review?	Rubric-based LLM evaluation, reliability analysis, Bayesian hierarchical modeling	Report · PDF · Project page
RAG Production Pipeline	How can retrieval, grounding, and monitoring improve RAG system design?	Hybrid retrieval, re-ranking, confidence calibration, observability design	Report · PDF · Project page

What to review

• Problem framing and evaluation design: Each report documents a defined problem, data/evaluation setup, and methodological choices.
• Statistical rigor: The projects use cross-validation, inter-rater reliability, confidence intervals, Bayesian inference, or calibration as appropriate to the task.
• Decision relevance: The work connects model results to practical review, triage, or monitoring decisions rather than treating a headline metric as sufficient on its own.
• Responsible use: Each package page states the limits of the project and the validation needed before any real-world use.

Selected results

Project	Reported result	Why it matters
AI Safety	96.8% classification accuracy; Krippendorff's α = 0.81	Separates annotation reliability from downstream classifier performance.
Breast Cancer	99.12% held-out accuracy; ROC-AUC 0.9987	Illustrates calibrated supervised-learning evaluation on the WDBC benchmark.
LLM Bias Detection	67,500 ratings; Krippendorff's α = 0.84	Demonstrates an uncertainty-aware workflow for large-scale content review.
RAG	94.2% citation precision; 96.3% Recall@10	Connects retrieval quality, grounding, and operational metrics in one systems design.

All figures above are reported in the linked technical documents. The AI Safety, LLM Bias Detection, and RAG results come from simulated evaluations built to demonstrate each method; they are not measurements of real models, publishers, or a deployed service. The Breast Cancer results use the public WDBC benchmark and are not independent clinical validation.

Technical focus

Python · SQL · R · scikit-learn · XGBoost · LightGBM · PyMC · ArviZ · FastAPI · MLflow · SHAP · OpenAI · Anthropic · Qdrant · Docker · Kubernetes

Repository guide

README.md                                      Portfolio overview
Resume_Derek_Lankeaux.md                       Résumé source
*_Report.md                                    Four technical reports
*_Publication.pdf                              Corresponding publication PDFs
project_packages/                              Per-project reader guides
generate_publication_pdfs.py                   Canonical PDF generator
requirements-pdf.txt                           PDF-generation dependencies
PDF_EXPORT.md                                  Build and validation instructions
scripts/validate_portfolio.py                  Artifact and local-link validation
.github/workflows/validate-portfolio.yml       GitHub Actions quality check

For PDF regeneration and validation, see PDF export instructions. The source notebooks referenced in the reports are not distributed in this repository; the reports and PDFs are the public portfolio artifacts.

Contact

Open to 2026 data science, applied ML, and LLM-evaluation opportunities. The best way to connect is on LinkedIn.