Introduction The last post ended on a word I had been circling for the whole series: trust. A calibrated score and a category label are only useful if the person reading them can see why the syste...

Introduction The last post ended by deferring a question: once the flywheel produces enough confirmed pairs to evaluate against, what does “it works” actually mean? I deferred it because the hones...

Introduction The last post argued that Vision Transformers, not CNNs, are the right architecture for postmortem dental imagery. The sharpest readers will have had one objection loaded before they ...

Introduction In system design, we talk endlessly about state-of-the-art models and benchmarks. But when you step away from the IDE and work as a medical examiner analyzing postmortem remains, you ...

Introduction In system design, we talk endlessly about edge cases and failure modes. But when you step away from the IDE and work as a night duty doctor in a geriatric hospital, you realize that h...

Introduction In the previous post, I argued that a search engine must come before the AI model, and I closed with a promise to get specific about the engineering. This post is that promise paid of...

Introduction In the previous post on representation learning, we discussed how to map complex dental data into computing spaces. However, before deploying flashy AI models, we must ask a fundament...

Introduction In the previous post, I discussed why forensic dental AI depends on more than just models. Before a system can learn anything useful, it must first have access to: Structured den...

Introduction In the previous post, I discussed why forensic dental identification cannot rely only on exact queries. A missing person identification system does not operate in a clean database en...

Introduction Most data systems are built around queries. Input a condition Retrieve matching records This paradigm works well when the data is complete and the queries are precise. Howeve...