AI Projects
Agent systems and automation I've built and actively run.
MyOps active
Personal automation and multi-agent orchestration system.
Problem: Running a home lab and small MSP-style environment means constant, repetitive ops work across AD/GPO, virtualization, backups, security monitoring, and ticketing — too much for one person to do by hand consistently.
Approach: MCP servers for Active Directory/GPO, Citrix, Veeam, Wazuh, Zammad, and network gear, wired into a Claude-driven task/ticket workflow that plans and executes ops work with a human checking the risky steps.
Multiagent web dev team (Loki Mode) active
The parallel multi-agent dev team that built this site.
Problem: Autonomous coding agents are useful but risky unsupervised — an agent with commit access can push broken or unreviewed code straight to a live site.
Approach: A parallel multi-agent dev team (Frontend/Backend/DevOps/QA/Reviewer roles) runs autonomously against a PRD, with a blind code-review gate before any PR merges and a human merge step before anything reaches production.
GhostWriter active
Local agent pipeline for video course production.
Problem: Producing a video lesson end to end — script, narration, visuals, and a rendered final cut — is a lot of manual, sequential work per lesson.
Approach: A local agent pipeline scripts a lesson, generates narration via TTS and visuals via Stable Diffusion, then assembles a rendered video automatically.
Budget Reconciler active
Monthly bank-vs-budget reconciliation with debt tracking.
Problem: Bank exports and budgeting-app exports drift apart over time, and manually cross-checking every transaction each month is tedious and error-prone.
Approach: Reconciles a bank export against a budgeting app's export, categorizes every transaction against a user-defined rule set, flags anything uncategorized or missing from either source, and tracks debt paydown over time. Rerun monthly.
LTX-Video experimental
Local video generation on local GPU hardware.
Problem: Cloud video-generation APIs are expensive and limit experimentation; iterating quickly on open video models needs local control.
Approach: Runs the open LTX-Video model locally on local GPU hardware rather than through a cloud API, for faster, cheaper iteration.