About
I build and study AI systems: local model inference, agent workflows, and tools that combine local compute with hosted models. This site documents my projects, experiments, and what I learn from running them.
My setup includes AMD GPUs, a Proxmox homelab, and applications built with TypeScript, Bun, and SQLite. I focus on model behavior, resource use, repeatable measurements, and making automated work easier to inspect.
Projects
Media Coordinator
A local AI workflow system for planning, image and video generation, and output review. Coordinates GPU stages, records each attempt, and checks the files produced by a run.
- Bun + TypeScript
- SQLite
- Local AI
- Job queues
Local AI on AMD
A two-GPU AMD system for local inference and model experiments. Measurements include workload, concurrency, and serving settings so results can be compared in context.
- AMD GPUs
- ROCm
- Local inference
- Measured experiments
Agent Arena + rig-wall
Local tools for inspecting agent coordination and system telemetry. Review a shared-board run, inspect a selected candidate, or replay recorded dashboard data. Included examples use synthetic inputs.
- Bun + TypeScript
- Shared board
- Telemetry
- Record and replay
Homelab
Self-hosted infrastructure for AI applications and experiments. Proxmox runs project containers, Tailscale connects the hosts, and systemd manages services. Monitoring and failure records help me understand how the system behaves.
- Proxmox
- Tailscale
- Cloudflare Tunnel
- systemd
Novel Assistant
An experiment in structured LLM workflows using long-form text as the workload. Plans, model output, and reviews are stored in files and Git. SQLite records model-call telemetry for inspection.
- Bun + TypeScript
- DeepSeek
- Files + Git
- SQLite telemetry
Work Interface
An operator interface for bounded agent work: preserve the exact candidate, inspect independent verification, and record a human decision. Semi-archived while development focuses on Media Coordinator.
- Bun + TypeScript
- SQLite
- Disposable VM sandbox
- Explicit review
What I'm exploring now
Local inference
Comparing model formats, serving engines, memory use, and concurrency on AMD GPUs.
Hybrid AI workflows
Exploring which tasks fit local models and where hosted models help, with attention to quality, latency, cost, and data boundaries.
Agent coordination
Testing bounded tasks, independent review, shared state, and explicit acceptance of agent output.
Evaluation and debugging
Keeping inputs, settings, outputs, and failure notes together to understand what changed between runs.