Reasonable Forge
Two brand sites live and the first product line in production, with its fits confirmed in print
The check Every part clears a printability rules gate before it's sliced, and a fit stays marked UNVERIFIED until it prints
LLMOps · evals · cost · observability
I build AI systems, and the checks that tell me whether they work.
I'm Ryan Walsh. I spent twenty years in regulated tax software and have been an Enrolled Agent since 2011. Now I design LLM systems where the eval, the cost ledger and the failure log are part of the build from the first commit.
Selected work
All projects → Agent Panel
Its enforcement chain is open source as the reasonable-agents plugin (MIT)
The check The vault agent exposes only typed-JSON intents, so private context can't leak into public output
Source on GitHub ↗Writing
All writing →- The Orchestrator Was Fine. The Limbs Misfired.
Evidence for the better-trained-limbs thesis: my content pipeline's orchestrator was solid; the limbs are where it broke.
- Better-Trained Limbs
A year and a half of drums and agent systems on the same learning curve. The goal turned out to be better-trained limbs, not a smarter orchestrator.
- I Was Wrong About the Advisor Pattern
I built an eval harness to check whether my advisor pattern was doing anything. It wasn't.
About
Claude writes the code. I design the system, read what it produces and check it against reality. I'm automation-first: if a task will repeat, the work is building the tool that does it. That habit is how I ended up in LLMOps, because you can't trust automation you can't measure.
More about me →