Eric Minish Personal site

About

I build agent and data infrastructure, and I came to it through operational systems rather than through research. Central Lake, Michigan.

The rule everything here follows

A system should decline to answer rather than guess. That sounds like a platitude until it costs something, so the work on this site is chosen partly to show what it costs.

It cost a designA retrieval engine I built beat keyword search on a small corpus and lost to it on a real one. The write-up leads with the loss, because a measurement that only ever confirms you is not a measurement.
It cost a resultA modelling target scored 38.8% better than its baseline and got cut anyway, once the right baseline showed it was a restatement of a variable already in hand.
It costs throughputThe compression proxy in front of my tool servers forwards the original payload untouched the moment it cannot prove a perfect round trip. Slower, and never wrong.
It costs convenienceMy knowledge base takes proposals from agents and reviews every one before it lands. Uncurated ingestion was the cheap alternative and it was rejected: at volume, the important things stop being findable.

How I got here

The professional work has been operational systems: workflow state, event handling, automation around CRM behaviour, and the reporting layer that has to make all of it legible afterwards. The part I care about is rarely the interface. It is the point where somebody needs to know what happened, what happens next, and whether they can trust it.

That turns out to be the whole problem in AI infrastructure too. An agent acts on state it cannot see directly, so the interesting work is the layer deciding what it is allowed to know, what it is allowed to write, and what happens when part of that goes wrong. Neither of those is a clever trick. Both are refusals to let a convenient default quietly become a correctness bug.

The homelab is the same argument in hardware. Ten machines carrying real routing, monitoring and recovery, so a claim about how something behaves under failure is something I have watched rather than read.

Email is the best way to reach me — [email protected].