I build machine learning models and the products that ship them.

Production client systems, applied-ML projects, and live in-browser demos. Public work runs in-page or links out; protected client work is previewable on synthetic data, and source and sensitive details stay private.

• 

  01

  Graduate projectLive

  KickCast

  Turns a 21,371-match feature matrix into 10,000-run Monte-Carlo odds for every one of the 104 games of the 2026 World Cup: full probability distributions, not a single guess.

  An end-to-end ML system that predicts international football matches as a 3-class (home/draw/away) problem over a 21,371-match feature matrix. From those probabilities it Monte-Carlo simulates the entire 48-team, 104-game 2026 World Cup. Seven classifiers plus a custom PyTorch net, SHAP explainability, and a live dashboard.

  • Python
  • scikit-learn
  • XGBoost
  • LightGBM
  • PyTorch
  • Optuna

  • 21,371 matches × 38 feat.Feature matrix
  • 7 classifiers + PyTorch netModels
  • 10,000 iterationsMonte Carlo

  Open live↗View source↗
• 

  02

  Protected

  Bastion

  Compresses a 10–12-week security engagement into one workflow: a multi-stage Claude pipeline (~80% token reduction) where every finding cites the evidence it came from and a human signs off.

  A production multi-tenant SaaS that runs the full cybersecurity-maturity assessment lifecycle (evidence vault, a multi-stage Claude gap-analysis pipeline against frameworks like NIST CSF 2.0, scoring, and client-ready reports) with a strict RLS-enforced consultant/client split. Live as an invite-only deployment, with a walkthrough available on request.

  • Next.js 15
  • React 19
  • TypeScript
  • Supabase
  • Claude (Sonnet + Haiku)
  • Pinecone (RAG)

  • 5Frameworks · NIST / CIS / ISO / SOC 2 / CMMC
  • cites evidence + human sign-offEvery finding
  • Supabase RLS (client/internal)Tenant isolation

  Private · request access
• 

  03

  Graduate projectCompleted

  INTELIPS

  Auto-labels 25,640 Enron emails with an LLM, then benchmarks five architectures. A weighted ensemble tops out at 72.85% macro-F1 on the real annotated data.

  A graduate NLP project that auto-labels ~25,640 Enron emails into Low/Normal/Critical priority with an LLM served via the Groq API, then trains and benchmarks five architectures: TF-IDF + metadata XGBoost, a context-aware MLP, a BERT + attention model, fine-tuned BERT, and a personalized user-embedding net.

  • Python
  • PyTorch
  • scikit-learn
  • XGBoost
  • BERT / Transformers
  • Groq API (LLM)

  • 72.85% (ensemble)Best real-data F1
  • 72.18% macro-F1XGBoost baseline
  • 25,640 (Enron)LLM-annotated emails

  View source↗
• 

  04

  Protected

  PayStream

  A full-stack payroll-automation platform built for multinational FMCG/pharma clients: a FastAPI + pandas/openpyxl backend that turns vendor payroll exports into bank payment files, GL/variance reports, and accruals, with a React config UI, JWT auth, and Docker deployment.

  • React
  • TypeScript
  • Vite
  • FastAPI
  • Python
  • pandas

  • bank files · GL/variance · accrualsOutputs
  • multi-currency · multi-DBPlatform
  • JWT/bcrypt · RBAC · audit logSecurity

  Private · request access

The full inventory behind the curated work above: 50 repositories, 8 public.

Two interactive toys I wrote from scratch (no ML library), running in your browser. Train a small neural network and watch its decision boundary form, or race four optimizers down a loss surface. Real code, not screenshots.