Zero-based redesign
Notes on rebuilding AI systems from zero.
The record
A company's software already records how it operates: orders, routing decisions, machine readings, the moments a person corrects the model. Taken together that record describes the business more accurately than any process document, because a process document says how the work is supposed to happen and the record says how it did. In most companies nobody reads it.
The hard part of integrating AI is rarely collecting data or picking a model. It is connecting what already exists and making it mean the same thing everywhere.
Corrections
When a person overrides an AI system, they produce the most useful record in it: a specific case where the system was wrong, with the correct answer next to it. Grouped over time, those overrides show which rules have drifted and which decisions the model should never have been handed. Most systems store them and query them for nothing.
Definitions
Accuracy on a real business problem usually turns on definitions, not model size. A few pages describing what a business means by its own terms move results further than a larger model or a full knowledge graph. Grounding helps most where data is thin and can hurt where it is dense, which is why a generic demo impresses and then fails on the specific case.
Governance
The properties that keep an integration safe, read-only access to source systems, a store the organization owns, changes that can be reversed and measured, are not a phase of the work. They hold across every stage or they hold nowhere. Treated as a separate step, they end up decorative.
An open problem
One problem here is unsolved. Automated judgment of generated visuals works on a still frame, and a still frame carries almost no information about motion. Two animations that differ only in timing produce nearly identical frames at any instant, so a frame-based score cannot tell them apart, even when one reads as correct and the other as broken.
Judging motion needs a measurement over time, not a snapshot. No cheap one exists yet. It's the problem I'm working on now.
Work
recall-eval is a zero-based rebuild of retrieval evaluation: agent memory retrieval that logs the component scores of every decision, so the thresholds can be measured against judged traffic instead of guessed. Source on GitHub.
motion-blind is the benchmark behind the note above: timing-only variants of an animation are indistinguishable from a single frame, at chance even for a production embedder like SigLIP, and perfectly separable from the ordered sequence. Source on GitHub.