This page is about the line between what AI can and cannot do. Models are unbeatable at recombining everything that has already been written; they cannot collect what has never been recorded, and they do not know which answer matters. Most of my work now sits on that line: defining what a system should do, identifying where it fails, and deciding what to build, what to fix, and what to leave out. The finished work lives at ai.html. This page tracks the part that isn’t figured out yet — and maybe never will be.
| Parent Directory | - |
| ai-product-artifacts/ | model + system behavior |
Work where the deliverable is the behavior of an AI system — what it says, what it withholds, and whether it should exist at all. It runs from a 2017 decision-tree triage chatbot grounded in Geek Squad repair data, through a concept test that told EY their recommendation engine was too simplistic for real client relationships, to HealthBot, a therapeutic chatbot rebuilt around a memory-suppression architecture. The recurring move is translation: clinicians, technicians, and benefits reviewers know their fields but are not programmers or product owners. This work turns what they know into what a product does.
| healthbot-suppression-architecture | defining what a therapeutic model withholds, not just what it says | |
| bestbuy-triage-2017 | delivering conversational self-diagnosis before LLMs existed | |
| ey-recommender-verdict | identifying an AI product that should not have been built | |
| veterans-claims-feedback-loop | recommending an architecture where outcomes teach the model |
| human-machine-translation/ | capability boundary experiments |
Small tools invented at the exact spot where the machine stopped understanding. When an AI could not read spatial intent, I built a plot-point coordinate language so an animation could be directed instead of described. When no tool could reproduce the phosphor glow of vector-era arcade games, I rebuilt the CRT effect itself. When notebook logic could not reach a development team, I packaged it so it could. None of these are client engagements. They are evidence of one skill: when the interface to the machine fails, change the interface until it works.
| animation-plot-points | defining a coordinate language for motion the model could follow | |
| crt-effect-rebuild | recreating the vector glow of asteroids, battlezone, tempest | |
| ml-notebook-handoffs | delivering notebook logic in a form developers can build from |
| field-research-usability/ | the input side |
The input side. A model can only recombine what has been recorded; fieldwork is the act of recording it. Twelve hundred data points from forty hours of EY interviews surfaced an incentive problem no document contained. Research operations at DroneUp were assembled from chat logs, employee testing, and borrowed marketing studies when there was no budget for any of it. At Intuitive Surgical, the finding that mattered — surgeons training each other in rogue Facebook and YouTube groups — existed nowhere in writing until someone went and found it. This section is why the other two are worth anything.
| ey-1200-datapoints | identifying the incentive problem behind a failed CRM | |
| droneup-research-ops | delivering research operations from zero, with no budget | |
| intuitive-rogue-communities | identifying the workaround communities off the official map | |
| chs-field-enrollment | identifying the half of 11,000 employees the intranet ignored |