I'm a long-time developer in the media industry, now focused on how AI can support decision-making and knowledge work. I build tools to power my own explorations, and open-source the components that might be useful to other people.
I'm building a media software factory and exploring how cutting-edge tools can make new forms and formats possible. Three active, public-facing experiments:
- NJ School Data - https://njschooldata.fyi - a collection of data and visualizations to help parents and other stakeholders navigate New Jersey schools. Largely driven and maintained by agents, with human editorial oversight.
- Backfield - https://backfield.net - An agentic news organization covering the AI & Media beat. Agent reporters maintain their own beats, write social media-style posts, collaborate on shared, durable artifacts, and evolve their own enterprise reporting.
- Caz Games - https://caz.lyra-forge.com - An agentic game studio, learning in public on how to make games that don't suck.
The vast majority of the code here is generated by AI. A lot of it is alpha quality. I'm making it available because it might work for your use case, or it might spark an idea for your tooling/agent to implement on your own.
| Project | What it is |
|---|---|
| somm | Self-hosted LLM telemetry, routing, and model memory. Run your codebase's LLM workloads through somm and it records calls locally, juggles access across different models and providers, and evaluates workload↔model fit so you're not overpaying. |
| spindle | Blend skills — composes source skills into surface-specific sets: resolve the right subset per repo, lint the blend for coherence, render per harness and model, and materialize it where the agent loads it. |
| Project | What it is |
|---|---|
| flip | Reporter's notebooks — git-friendly research corpora maintained by humans and agents in the same files. Sources are hashed at capture, claims are gated by a corroboration bar, and LLM output is a lead — not evidence — until promoted. A wiki tells an agent what we know; a notebook can prove where it came from. |
| eno | Vault intelligence — structural gardening for a knowledge store like an Obsidian vault, so you and your tools can both work in it. CLI, MCP, and HTTP; light core, local-first. |
| Project | What it is |
|---|---|
| artoo | Artifacts — self-contained HTML mini-sites that pair a presentation with the research behind it. Renders from file://, vendors its assets, deploys to Pages/rsync/anywhere, and the research directory can never accidentally ship. |
| artoo-mermaid | Offline diagrams for artoo — pinned Mermaid vendored into artoo artifacts, so diagrams render from local files, forever. |
| ergo | Data pages — dataset documentation as datasets actually are. The real burden isn't the schema, it's the caveats: every known issue gets a stable ID, a type, a machine-readable scope, and a link to the code that works around it. Data + documented caveats ⇒ justified use. |
| vizier | Chart decisions and critique for journalistic data visualization. Deterministic chart-form recommendation and colorblind-safe color math (no LLM, no keys), plus corpus-backed critique of finished charts. What it suggests is what it would pass. MCP-native. |
- des — a design engineer driving consistency for lyra forge tools and products
- backfield-client-sdk — zero-dependency Python SDK for backfield.net, a human/agent-blended space where agents participate as first-class citizens as long as they're legible, governed, and answerable to a named human