Skip to content

Policy Driven Cryptographic Agility

    Project Summary: Policy-Driven Cryptographic Agility & Data Security Decision Engine

    This repository contains two related governance systems, built by the same design philosophy but kept deliberately separate:

    1. Cryptographic Policy Engine (R/00–R/10)
    Maintains an auditable inventory of where cryptography is used across an organization’s systems, a version-controlled policy defining which algorithms are currently approved (including post-quantum and hybrid algorithms), and an engine that evaluates and durably logs which algorithm applies to which system — reading inventory facts itself rather than trusting caller claims.

    2. Data Security Decision Engine (R/11–R/14)
    A correlation-aware, multi-signal risk engine that recommends ALLOW / WARN / BLOCK for data-movement events (e.g., sensitive data headed to an external destination). It evaluates signals against a governed policy table and discounts correlated signal pairs rather than naively summing them, producing a durable evidence trail for human security/compliance review.

    Shared design philosophy:

    Don’t trust a single signal — gather evidence from the systems that actually own it
    Control explicitly for correlated or contradictory evidence rather than combining it naively
    Evaluate against an external, versioned, content-hashed policy — never an ad hoc decision
    Make a defensible, attributable, durably-logged decision
    Stop short of enforcement and stop short of the final human judgment call

    What it is not: neither engine is a cryptographic implementation or an enforcement point. Real deployments still need HSMs/PKI/validated crypto libraries for the first system, and real DLP/IAM/CASB enforcement for the second — this project governs the decisions, not the enforcement.

    Structure: R scripts (R/), CSV-based config and policy tables (config/), tests (tests/), an R Markdown report (report.Rmd) that renders the live inventory/policy state, and a full worked example comparing both engines side by side.

    Breast Cancer Diagnostics

      This project applies supervised machine learning to predict whether a breast tumor is benign (B) or malignant (M) using the Breast Cancer Wisconsin (Diagnostic) dataset from the UCI Machine Learning Repository.

      Highlights:
      Ensemble model achieved over 99% accuracy on the test set.
      – Perfect sensitivity (no malignant cases misclassified).
      – Most predictive features: Concave Points (Worst), Radius (Worst), Perimeter (Worst).

      Demonstrates how ensemble learning methods can enhance diagnostic precision and reduce false negatives in medical classification tasks.

      Employment Forecasting

        Every month the U.S. Bureau of Labor Statistics publishes employment figures for large sections of the U.S. Economy. The task is to forecast the amount of people employed in the financial activities sector for four years. This information is publicly reported by the St. Louis FRED and can be found at the following link. https://fred.stlouisfed.org/series/USFIRE.

        Four years represents a large amount of time and confidence windows will be wide. Therefore, I will test a few forecasting models by comparing their respective RMSE and Winkler scores on a known portion of the time series.

        Computer Vision System: Dry Beans

          This project is a computer vision–driven multi-class classification system🚀 It distinguishes seven varieties of dry beans using image-derived features from the UCI Machine Learning Repository (13,611 samples, 16 features). The best-performing model was a Radial Support Vector Machine, achieving ~93.5% accuracy on the hold-out test set, outperforming the ~93.1% benchmark reported in prior research using SVMs.
          The results suggest further gains could be possible with additional, less-correlated features (e.g., color information from raw images)

          This project applied end-to-end machine learning techniques and demonstrated how careful preprocessing, model selection, and validation can meaningfully improve real-world classification performance.

          MovieLens

            Recommendation systems have revolutionized e-commerce. Companies like Amazon sell all manner of products to their customers and actively collect reviews which are stored in massive databases. Machine learning tools are then used to recommend products across the full range of customers in a tailored and highly specific way. This process enables data driven companies to increase their pool of satisfied, returning consumers. It is the same process used by media companies like Apple Music and Netflix.

            This project will use a large subset of the MovieLens data which holds approximately 10 million ratings on approximately 10,000 movies from around 70,000 users. I will use the Root Mean Squared Error of my predicted ratings against a validation set of actual ratings in order to evaluate several different models. The goal is an RMSE of below .86490 (which corresponds to within one star of accuracy on a Netflix type rating system).