Local AI Diet Coach
My own AI agent for personal nutrition support, combining local inference and persistent knowledge. An open-source release is planned.
From an everyday question to a personal system
This project began with using ChatGPT during a diet: looking at a meal, adapting a recipe or finding something to cook with available ingredients. It also raised a question: could that personal nutrition and health context stay in my own environment?
That became a local AI Diet Coach. The public PFTinfo account dated 12 August 2026 is the basis for this page. Current implementation details and development status still need to be updated from the project itself.
A purpose-built agent
I built the Diet Coach agent for my own daily workflow. Like the agents in ARtBase, it is an exploration of the tools, context and memory that make a model useful within a larger system.
Planned open-source release
Diet Coach is intended for publication as open-source software. This site will provide its public project page and links to the repository, documentation and release material when available. The repository URL, license and release date have not yet been announced.
The model should be replaceable
The central architectural choice is that knowledge persists when the model changes. The agent uses personal context, diet rules, daily logs, recipes and earlier feedback. The language model is a reasoning layer inside the application.
The source describes a locally running Gemma model. The exact version, hardware configuration and repository have not yet been established for this page.
Vision with room for corrections
A meal photograph can provide a useful starting point. It remains an estimate: portions, hidden ingredients and preparation methods are not always visible.
Original analyses and later corrections are therefore kept distinct. Repeated feedback can suggest a hypothesis without turning every individual correction into a general rule.
Kitchen Lab
Recipes can have multiple versions. Adjustments and experiences are preserved so that the next attempt builds on earlier results. AI becomes a tool within an iterative process.
Local processing and explicit uncertainty
According to the public project account, personal data and inference remain in my own environment. The system distinguishes observations, experiences, hypotheses and confirmed knowledge.
Diet Coach is a personal experiment, not a validated medical device. The source discusses a possible limited clinician view as a future idea; this page does not present it as an existing feature.
Still to document
- Current architecture and model versions.
- Hardware and bounded inference configuration.
- Screenshots without personal health data.
- Any public code, installation instructions and evaluations.
See also my perspective on AI in healthcare and pulmonary function.