The dated record of what I have actually built. A technology generalist who ships end to end: AI agents, real software, data pipelines, security tooling, investigative infrastructure. Not a list of employers. A list of things that exist because I built them.
Newest first. Click any project to open the full stack and story.
Agents that reason about truth, a chief-of-staff layer, and real business bets.
PurposeMoves photos from a phone straight into the laptop workflow with one step, killing a small daily friction.
NotableSmall and sharp. The kind of friction-remover that quietly makes everything else faster.
PurposeThe single source of truth and operating brain for Maverick Logistics LLC, a diesel-fuel delivery carrier in Salem, South Carolina. Designed to grow from a knowledge base into an autonomous operating agent for the business.
NotableNot a toy. A real company's operating brain, 102 commits deep, built to run on its own server.
PurposeWhen an Android phone starts forcing gibberish into every message (you type "lamb", it types "lambbgfg"), this resets the keyboard's corrupted memory and clears it up in about a minute.
NotableThe range in one repo: from a twelve-agent fleet down to fixing a phone over a USB debug cable.
PurposeA coordinated "company" of purpose-built agents, each with its own lane, personality, and playbook: Czar (business development and outreach), Quill (writing), Parallax (the epistemic layer, credence and prediction), Jethro (job search), and Tutor (learning).
NotableNot one big AI, a division of labor. Specialists that hand work to each other under a human gate.
PurposeA proactive, never-command agent that studies all the repos, goals, and working style, asks questions, learns continuously, and runs constantly-shifting triage to surface the single most important task first, under strict security discipline.
NotableAn orchestrator for the orchestrators. (The agent that built this website, CAIRN, is one of GIDEON's domain subagents.)
PurposeAn iZurvive-style interactive Chernarus map you can annotate with your own bases, stashes, death spots, and loot runs, synced live between two players. Built to grow into kill logs, a war journal, and loadouts.
NotableReal-time two-player state sync on a custom Leaflet map. A full-stack app that happens to be about a video game.
PurposeA round-the-clock research fleet built to check claims, not just generate them. Cheap models do the high-volume scanning; a Claude reasoning layer judges only the hard cases; the human keeps the final say.
NotableCost-tiered by design: pennies on the reading, judgment only where it matters. Verification over generation.
PurposeSnap a photo to identify a plant, get a weather-aware watering schedule, and let a proactive intelligence layer surface problems early. Built as an installable PWA, designed to grow into a social and marketplace platform.
NotablePhoto identification plus weather-aware logic plus a proactive layer, a real product, not a to-do list with a plant theme.
Agents that talk to each other, and the servers that run them.
PurposeThe evidence backbone for investigative journalism and publishing. A service (in the same category as JayBrain, not an agent): the research agents call into it to store entities, sources, claims, and citations.
NotableSchema design and evidence modeling for reporting on third parties. The foundation under a Bellingcat-style publishing play.
PurposeA task dispatcher with a registry and audit logging that hands work across the agent fleet and keeps a record of who did what.
NotableThe piece that turns a pile of separate agents into one coordinated system.
PurposeA real-time, voice-driven Central Command interface, single-tenant and cloud-only, wired to a mission-control dashboard.
NotableA self-hosted real-time voice stack, open-source LiveKit and local TTS instead of a paid API, on its own server at spirality.dev.
PurposePolls podcast RSS feeds, fetches and classifies episodes, and runs tiered transcript orchestration, feeding the downstream research stack (Parallax, Quill, Dossier). Uses Siphon to fetch and extract, persists to Dossier.
NotableA real ETL pipeline with daily polling and tiered transcript handling, one clean stage in a larger research machine.
PurposeNamed for the shield of Athena: it protects, it does not build. Aegis watches the whole ecosystem, catches failures before they cascade, monitors resource limits, and oversees the agent fleet so the rest can keep running.
NotableThe always-on backbone. 170 commits of the unglamorous reliability layer that lets a one-person fleet run unattended.
The systems everything else stands on.
PurposeA production knowledge-management system for the Ellsworth service business, built on a three-tier architecture: a Salesforce mirror (Tier 1), a canonical markdown vault (Tier 2), and downstream consumers including dashboards and equipment tracking (Tier 3). Runs on a live server.
NotableBuilt over 12 weeks and 4 phases. 450 commits, the most-worked repository in the whole portfolio, and it runs a real business.
PurposeTook a sprawl of working-but-unmanaged sub-agents and tools and brought them under version control at once: kai (scheduling and coaching), echo, sieve (content and transcript extraction), watchtower (monitoring), youtube-transcriber, signalforge, successforge, and kit.
NotableThe discipline of knowing when to stop hacking and formalize. A one-day cleanup that turned chaos into a managed portfolio.
PurposeA field-ready equipment survey app for the Ellsworth job, installable to a phone and usable on site. 51 tracked files. The direct precursor to Elmer.
NotableA real tool for a real operational problem, the first version of what became a 450-commit business system.
PurposeA production-grade curriculum for achieving complete mastery of Claude Code: 27 modules across 5 tiers, with a "vibe coding first" approach and a 95 percent mastery target.
NotableYou do not write a 27-module course on a tool you only half understand.
PurposeCustom Sigma detection rules developed and tested in a Windows Active Directory homelab, each validated against real attack simulations with Splunk Enterprise as the SIEM.
NotableIncludes a Kerberoasting RC4 TGS detection (ATT&CK T1558.003) written from an actual attack run, not copied from a template.
PurposeAn MCP server that gives Claude persistent memory, user profiling, task tracking, spaced-repetition learning, and cross-session continuity, the memory layer behind the entire fleet.
NotableHybrid vector-plus-keyword search with locally-run embeddings, no cloud vector database. People pay a monthly fee for a worse version. 227 commits.
The roots. The tools everything after was built on top of.
PurposeA Windows Active Directory lab (Server 2022, lab.local) with Splunk as the SIEM, built to simulate real attacks and write detections against them. The environment behind the sigma-detection-rules work.
NotableReal blue-team infrastructure, attack simulation and detection, not a tutorial follow-along.
PurposeA comprehensive scraping framework that automatically picks the optimal strategy per site: a fast decision tree routes each target between plain HTTP, a full Playwright browser, or a browser extension, based on the site's characteristics.
NotableThe interesting part is the decision engine, it classifies a site and routes to the cheapest method that will actually work.
PurposeCrawls a site and extracts text, images, and structured data (JSON-LD, Microdata, RDFa, OpenGraph), with optional headless-browser rendering for JS-heavy pages. Still the canonical fetch-and-extract layer behind newer tools today.
NotableThe earliest project on record, and it still ships work. Stealth rendering means it handles bot-guarded and JS-heavy sites, not just static HTML.