Projects

Three things I’ve built. Each one starts with a problem, runs on real code, and ends on what the evidence actually said.

Local Offline AI Assistant

A privacy-first assistant that runs entirely on-device. It handles chat, retrieval, voice and disk analysis on quantized models, on an 8 GB laptop, with zero runtime network calls.

  • The privacy guarantee is enforced in code, not promised on a page. A policy layer refuses any cloud-hosted model, and every tool is gated behind a risk tier with recorded consent and an audit ledger.
  • 52 modules and 286 automated tests. An eval harness scores retrieval and generation separately, and fails the build when either one slips.

A hard no-network rule didn't cripple the assistant. It forced better engineering.

Python · Ollama · FAISS · ONNX embeddings · faster-whisper · Piper · SQLite · pytest

Model Drift & Data Monitoring

A monitoring pipeline for models in production. It compares incoming data against the training baseline and raises a flag when too many features have drifted, closing the post-deployment gap most ML projects skip.

  • Two independent signals per feature: a relative mean shift, and a two-sample Kolmogorov–Smirnov test.

Two cheap signals beat one clever one. When the failure is silent, coverage matters more than sophistication.

Python · pandas · NumPy · SciPy · pytest · GitHub Actions

Hybrid Network Intrusion Detection

A two-stage intrusion detector on the KDD Cup 1999 benchmark: 494K connections, 41 features. An unsupervised Isolation Forest flags anomalies, then a supervised Random Forest names the attack family.

  • Evaluated on a stratified split with per-class recall and macro-F1, rather than the accuracy number that 79% class imbalance would have flattered.
  • The cascade lost. It scored a macro-F1 of 0.73 against 0.95 for a plain Random Forest. Stage one's recall collapsed on the rarest attack classes, and no later stage can recover what stage one never passes on.

I kept it in anyway. A design that fails on the record beats one that fails quietly.

Python · scikit-learn · pandas · NumPy · Jupyter

The code for these lives on github.com/blzee-maker.