Stdlib-only guards and a falsifiable intent-record system that let an LLM coding agent run unattended without trusting it to bless its own work — mechanical, testable, and enforced outside the model.
- Python
- AI safety
- Claude Code hooks
- Agent Skills
Read the write-up: excubitor — Safety Fences for Autonomous Coding Agents →Code for excubitor — Safety Fences for Autonomous Coding Agents ↗ (opens in a new tab)
A deterministic tool that reconciles an AI coding agent's natural-language claims — 'the tests pass', 'I created the file' — against what its Claude Code session actually did, calibrated to zero false accusations across 400 real sessions.
- Python
- AI safety
- CLI
- Static analysis
Read the write-up: did-it — Checking an AI Agent's Claims Against Its Transcript →Code for did-it — Checking an AI Agent's Claims Against Its Transcript ↗ (opens in a new tab)
A personal project in development: Python orchestration, MCP execution, React, and Rust/Tauri, with permissions separated from prompts and explicit recovery safeguards. Architected and directed with AI-assisted implementation.
- Python
- MCP
- React
- Rust / Tauri
- Kotlin
- Evaluation
Read the write-up: Gestalt — A Multi-Provider AI Workspace →
A locally running, credentialed Python utility used by colleagues to work with Oracle databases through Claude and Cursor.
- Python
- Oracle Database
- PL/SQL
- Claude
- Cursor
Read the write-up: OracleDB AI Bridge — An Internal Developer Tool →
A pre-LLM voice-assistant spike assembling the speech-and-vision stack of its day — Google STT, neural punctuation restoration, nltk parsing, pyttsx3 speech, and an OpenCV face pipeline — built before any of it was a single API call.
- Python
- speech_recognition
- pyttsx3
- nltk
- OpenCV
- Punctuator (BRNN)
Read the write-up: Walter — A Pre-LLM Voice Assistant →
An interactive touchscreen map for a septic-services business: a customer spreadsheet reverse-geocoded into a filterable Folium map with pricing zones, service-due heatmaps, and upsell alerts — on a repurposed big-screen TV with an infrared touch overlay.
Read the write-up: BAT Powerwall — Geospatial Service-Mapping System →
An AI red-teaming experiment: one GPT-4 jailbroken by another to generate novel malware from real samples — a demonstration of model-guardrail failure, never released. Credited at SANS HackFest 2023.
- GPT-4
- AI red-teaming
- Jailbreak research
- AI safety
Read the write-up: HackGPT — An AI Red-Teaming Experiment →
Historical voice-assistant development with wake-word detection, speech recognition, pluggable LLMs, MCP tools, and speech synthesis. Recovery and repair are required before an end-to-end demonstration. Built AI-assisted.
- Python
- openWakeWord
- Silero VAD
- faster-whisper
- Ollama / Claude
- MCP
- XTTS
Read the write-up: Jarvis — A Voice Assistant Under Recovery →
A fully on-device, FOSS pipeline that fine-tunes a custom text-to-speech voice from a handful of audio samples — decode, loudness-normalize, segment, transcribe, fine-tune, synthesize, evaluate. A separate model-training workflow designed for integration with Jarvis. Built AI-assisted.
- Python
- Coqui XTTS-v2
- faster-whisper
- Audio DSP (LUFS, VAD)
- PyTorch
- Poetry
Read the write-up: voice-training — A Local TTS Fine-Tuning Pipeline →
A personal experiment comparing coding-agent results against frozen acceptance checks, with explicit limits on what a small provider comparison can establish. Built AI-assisted.
- Python
- Coding agents
- Evaluation
- Acceptance testing
Read the write-up: Coding-Agent Evaluation Harness →
An experimental Python router for portable prompt guidance, plus controlled comparisons that retained a negative result. AI-assisted personal development.
- Python
- Prompt guidance
- Evaluation
- Context budgets
Read the write-up: LLM Guidance Routing and Evaluation →
A self-managed Debian homelab — GitLab, Vaultwarden, Grafana, Docker behind a reverse proxy — plus a from-scratch web and mail stack for a dozen-plus domains. Most recently, a fleet-wide email-auth remediation that fixed outbound mail silently failing at the major providers.
- Debian
- Docker
- GitLab
- Grafana
- nginx
- Postfix
- SPF / DKIM / DMARC
- DNS
Read the write-up: Self-Hosted Infrastructure & Email Deliverability →
Packaged scientific-Python tooling behind my NASA and UMBC atmospheric research — crawlers, readers, and regridders that turn raw satellite and observatory data into analysis-ready netCDF.
- Python
- netCDF4
- HDF4 (pyhdf)
- cartopy
- multiprocessing
Read the write-up: Earth-Science Data Tooling — Packaged Scientific Python →Code for Earth-Science Data Tooling — Packaged Scientific Python ↗ (opens in a new tab)
A Monte Carlo photon-transport simulation for cloud radiative transfer, built as a student and later used to grade peers' implementations as a teaching assistant.
Read the write-up: Monte Carlo Cloud Radiative-Transfer Simulation →
A published converter that reads legacy HDF-EOS2 satellite granules and writes CF-annotated netCDF4 with the geographic coordinates reconstructed, so decades-old data opens in current tools.
- Python
- netCDF4
- HDF-EOS2 (pyhdf)
- pyproj
- PyPI
Read the write-up: ncarnate — Legacy Scientific Data to Modern netCDF4 →Code for ncarnate — Legacy Scientific Data to Modern netCDF4 ↗ (opens in a new tab)
A one-command file-encryption tool and library on PyPI — password-derived keys, an authenticated cipher, and a versioned file format, carried through the full lifecycle of a published package.
- Python
- Cryptography
- CLI
- PyPI
Read the write-up: pydlock — A File-Encryption Package on PyPI →Code for pydlock — A File-Encryption Package on PyPI ↗ (opens in a new tab)
An open-source Adobe Brackets plugin adding syntax highlighting and comment shortcuts for 19 languages the editor didn't support out of the box — from COBOL and Fortran to Julia, R, and Swift — via CodeMirror modes.
- JavaScript
- CodeMirror
- Adobe Brackets
Read the write-up: More Languages for Adobe Brackets — Editor Syntax Plugin →Code for More Languages for Adobe Brackets — Editor Syntax Plugin ↗ (opens in a new tab)