← Work · Multi-agent AI
Nutriscan
A small system of cooperating agents that turns a snapshot of a grocery shelf into a quiet, considered nutritional summary, designed for parents and caretakers who would rather understand than be told.
The question
Nutrition labels are written for the literate, the unhurried, and the able-bodied. They assume good eyesight, working memory, and a tolerant relationship with small print. Nutriscan begins with a simpler question. If a parent could ask a calm assistant for help understanding what is on a shelf, what would the assistant actually need to do?
The shape of the system
The work is structured as a small ensemble of agents, each with a narrow responsibility, and a coordinator that holds the conversation together. The intent is restraint rather than ambition. Each agent does one thing the way a thoughtful person might do it, and the coordinator is the one who knows when to ask for clarification and when to fall quiet.
[placeholder] Final agent roster and protocol diagram will be added once the system passes its second evaluation round.
Four agents, one coordinator
A vision agent reads the photograph. A label parser separates ingredients from nutrition facts. A reference agent compares the result to dietary guidance the user has previously shared. A presentation agent shapes the response into something a person can read while a child is asking a question. The coordinator asks each agent only what is needed and stops when an answer is honest enough to stand alone.
Accessibility decisions, built in
A photograph is a high barrier to entry for many users. The interface accepts a typed product name as a peer to the camera input. Returned summaries are written at an approachable reading level by default. Numerical detail is available on request rather than imposed. The response is structured so a screen reader can land on the most relevant sentence first.
The ambition is not a smarter scanner. It is a system that knows when to be quiet, and when a person would rather understand than be told.
Originality and safety
The project does not reproduce or wrap any existing nutrition application. The agent architecture, prompts, and interface are authored from a clean slate. Sample images used during development are taken in-house or sourced under permissive licenses, with documentation kept alongside the project. [placeholder] Final licensing notes will appear here once the dataset is locked.
Where it is going
A second round of evaluations is planned for spring 2026, with a small group of caretakers who manage household nutrition for family members with specific dietary needs. The case study will be updated as that work concludes.