Deployed experimental demo · Noncommercial
ChefAI
A recipe is only the beginning.
Cooking gets interesting when the instructions stop being enough. An ingredient is missing. A sauce breaks. A substitution changes the texture. Someone wants to understand why a technique works before deciding how to adapt it.
ChefAI explores what a specialist AI can do with those kinds of questions. It brings culinary knowledge, retrieval, reasoning, and structured cooking workflows into the same application.
The engineering challenge is connecting an answer to something useful. A conversation may lead to a recipe, a shopping list, preparation steps, or a guided cooking sequence. An explanation may need source material, a cooking mechanism, or an ingredient relationship—and those kinds of support are not interchangeable.
That distinction shapes the system. Flavor relationships can suggest possibilities. They do not automatically establish a factual claim. A generated explanation still needs to be distinguishable from the evidence used to support it.
ChefAI is a deployed experimental demo and a separate branch of the portfolio: a place to explore how these engineering ideas behave inside a specialized domain. I am keeping it noncommercial because of the particular resources used to build its knowledge base. Its value here is the system, the experiment, and the lessons from making domain knowledge usable.
An answer and its support are different things
The intended design separates source knowledge, cooking mechanisms and structured relationships, flavor information, and generated synthesis. Source records support factual claims; mechanisms explain processes; flavor relationships suggest possibilities. The model combines these into an answer, while recipe, shopping, prep, and cooking panels hold workflow state.
Enforcement does not yet consistently match that design. A recent offline review found unresolved intent-routing and guard-precedence defects, evidence labels that can be produced without the claimed retrieval, and support scoring that can count words in the question as evidence. These issues are not presented as repaired.
Recipe structure alone does not establish that every dietary, allergy, equipment, or timing constraint survived a workflow change. ChefAI remains an experiment, not a validated food-safety authority. Live model quality still needs separate evaluation.