User needs are more valuable than assumptions.

I’m Chris Hart. Accorderly is my consulting practice for small product teams. My role covers product, UX research, and development. I’m a freelance Chief Product Technology Officer (CPTO).

Useful knowledge-driven apps (AI or not) rest on three parts that work as one:

• Models parse external facts into a logical data format. Machine learning is good at pattern matching: it recognizes things it has seen before. Smaller and cheaper models are usually the best fit for embedding in applications. Sometimes there is no ML needed and a plain old search index works well.
• Rules check the processed data for correctness and implications. Rules and logic are good at explanations: they show why a decision follows from the facts. And they’re totally deterministic, which means control over potential liability.
• People drive the living goals and the desired outputs. A person reviews the result and can see how it was reached. But beforehand they also configure domain-specific rules that cannot be left to mere probability. This is a step above so-called “human in the loop”, and completely avoids the “reverse centaur” trap coined by Cory Doctorow.

It runs on proven data engineering pipelines that have powered enterprises for over a decade before the latest AI models were even released.

I have commercialized this AI framework with many companies including the ones shown above. These systems show their work, which is what makes AI safe to use where “the model said so” is not good enough.