On this page
Questions you could ask me
Use Up Food was one of my first complete attempts to turn generative AI into a working application. I built it around a familiar problem: you have some ingredients, you would like to use them, and you need a little inspiration for what to cook.
It was an experiment rather than the beginning of a serious commercial venture. That was deliberate. I wanted a subject I could understand well enough to question the answers, demonstrate to other people and improve by observing what happened.
Why food was a good place to start
Generative AI was developing quickly, and I wanted to understand what it meant to use it inside a product rather than simply type questions into a chat. Recipe generation offered an unusually accessible test. Almost anyone can understand a proposed meal, recognise a missing ingredient or spot an instruction that does not make sense.
That familiarity mattered. When an application is entirely new, it is easy to mistake an impressive demonstration for a useful result. With a recipe, I could examine the ingredients, method, order of operations and final presentation without needing a specialist to interpret the output.
A two-stage generation process
The application took an ingredient-led request and used AI to generate a structured recipe. It then used the resulting recipe to produce a related food photograph. The sequence mattered: the photograph needed to represent the dish the application had actually described, rather than a generic image loosely related to the original request.
I had to make choices about the format and presentation of both outputs. Recipes needed a consistent structure that a person could follow. The accompanying imagery needed to feel as though it belonged to the same product, even when the subject of the photograph changed from one meal to another.
There was also a practical constraint. Image generation took noticeably longer than text generation at the time. I limited retries to three attempts rather than allowing an open-ended process. That decision wasn’t an elegant optimisation; it was a response to the actual experience of using the technology then.
Consistency was harder than novelty
Producing one interesting recipe and one attractive photograph was relatively straightforward. Getting different results to follow a coherent format was much harder.
Each recipe needed recognisable sections and a readable sequence. Each photograph needed a similar visual language: lighting, framing, food styling and the impression of being photographed in the same studio. Prompting helped, but it did not make consistency automatic. The results came close rather than achieving perfect uniformity.
That distinction was useful. It showed me that a convincing individual AI result is not the same thing as a dependable application experience. The surrounding design, instructions, constraints and checks are what turn isolated generations into something that feels like a product.
A demonstration people immediately understood
One of the advantages of Use Up Food was how little explanation it required. I could show someone a few ingredients, let the application produce a recipe and related image, and discuss the result directly.
I demonstrated it to a PhD data scientist in Norway who had worked with machine learning for years. His reaction was enthusiastic. I remember that because it was encouraging, not because one person’s response constituted a formal evaluation of the application.
More importantly, the project gave me something concrete to discuss. I was no longer talking only about the potential of generative AI; I had a working example with visible strengths, weaknesses, compromises and design decisions. It also helped me demonstrate my developing practical experience when I moved into AI-related product work at Zegeba.
What I learned
Use Up Food taught me several lessons that have stayed with me. AI-generated content needs a defined role in the application. A process that works once needs repeatable structure if it is to work as a product. Image generation has a cost in time and variability. And the best way to understand those limitations is to build something, try it repeatedly and pay attention to the result.
There were no claims of production scale, commercial traction or independent validation. It was a completed learning project, and that is how I regard it today.
Its importance is in what followed. The early effort to control the relationship between a recipe, its presentation and a generated image was a small beginning of a much larger interest in directing AI-assisted work, questioning plausible outputs and keeping responsibility for the result with the person building the system.
I’ve written more about that development in 10,000 Hours of AI: What I’ve Actually Learned. Use Up Food is one of the earliest practical examples in that longer story.
Project context
Topic: AI-assisted engineering
Explore further
Further reading is listed below.