Sensory Sprout / 0→1 AI product system
Making new meals work for everyone at the table
Sensory Sprout learns from the meals a family already rotates, adapts recipes they want to try, and finds new ones, like banana bread, that everyone will eat.
Overview
An AI meal tool that adapts recipes to each family member’s food rules
- Problem
- Families couldn’t predict whether a meal would be accepted before it reached the table.
- Role
- Product Designer & Builder
- Research
- 6 interviews (5 adults, 1 child) and 40 test cases for the profile chat
- Where AI sits
- Interprets texture, preparation, and brand nuance. Families approve every profile change.
- Outcome
- Working prototype. Profile creation tested, recipe adaptation explored. Version 2 in progress.
01 Problem
Will this meal work tonight?
For families with sensory-sensitive kids, a new recipe is a gamble. It can match every listed restriction and still be refused because of its texture, its brand, or how it was cooked. Recipe tools filter by ingredients and nutrition, but that isn’t how these families decide what’s safe to serve.
02 Research & Synthesis
A matching ingredient could still fail at dinner
I interviewed five adults and one child about mealtime. Families described acceptance rules far more specific than any recipe filter could capture. A food was rarely accepted or refused on its own. It depended on how it was made, what brand it was, and what it touched.
Rules participants described
- Broccoli must be steamed, not roasted
- Peppers must be raw, not cooked
- Pasta is acceptable only with butter
- Foods cannot touch
- Brands are not interchangeable
In their words
I microwaved her plain Alfredo pasta and the sauce at the edge turned slightly brown. She wouldn’t eat it.
I know what’s in the frozen burrito. With other burritos, sometimes you don’t know what’s in them.
03 The pivot
Families didn’t need more meal ideas. They needed to protect the ones that worked.
Meal discovery Acceptance reliability
Most households already had a small set of meals that consistently worked. Their harder problem was knowing whether a new meal would work, and keeping trusted foods from disappearing through repetition, burnout, or small changes in preparation. Protecting a meal meant capturing the details that made it work, like steamed broccoli instead of roasted.
04 Product Model
Three questions every recommendation had to answer
Once safe foods became the center of gravity, recommendation quality was not enough.
Evaluate
Will this meal work tonight?
Check it against household acceptance conditions before cooking.
Protect
What already works?
Preserve safe foods, prep rules, and routines.
Expand
What’s the lowest-risk next step?
Suggest small changes from accepted conditions.
05 Product Decisions
The model learned slowly by design
Once the goal became protecting safe foods, the challenge shifted from recommendation quality to trust. Three decisions shaped how the system earned it: show the data behind every recommendation, learn from real meals, and let families approve every change.
View the early prototype
Decision A
Show the data behind every recommendation
People trust a recommendation when they can trace it back to what they told the system.
- Problem
- Families needed confidence before investing effort.
- Decision
- Each recipe shows which profiles, restrictions, and preferences shaped it, and flags conflicts before cooking.
- Tradeoff
- The experience is more complex than a simple recommendation feed.
- Intended impact
- Families can inspect, challenge, or adapt a meal before cooking.
Key
- Household context — Checked against active profiles.
- Profile notes — The rules the recipe was checked against.
- Acceptance risks — Conflicts flagged before cooking.
Decision B
Learn from real meals, not predictions
What a family says at the table is more reliable than what a model predicts in advance.
- Problem
- Plausible recommendations could still fail at the table.
- Decision
- After each meal, families record whether each person ate it, tried it, or refused it, and why.
- Tradeoff
- Learning is slower, because it depends on feedback.
- Intended impact
- Recommendations are grounded in what actually happens at the table, not assumptions.
Key
- Actual outcome — What happened at the table.
- Failure or success context — Why the meal worked or failed.
- Evidence trail — Saved before the system learns.
Decision C
Let families approve every profile change
The family, not the model, decides what becomes a household rule.
- Problem
- If the system learned automatically, one meal could rewrite a household preference.
- Decision
- The system suggests a profile note from meal feedback, and families save, edit, or dismiss it.
- Tradeoff
- The system learns more slowly.
- Intended impact
- Families keep control over what becomes part of the profile.
Key
- What it noticed — The meal outcome behind it.
- Suggested change — The exact note, before saving.
- Family approval — Saved only when the family confirms.
06 AI Architecture
Families set the boundaries. AI works inside them.
AI handles judgment, like interpreting preferences and adapting recipes. Families approve everything that’s saved. Testing showed safety rules can’t live in a prompt, so the next build moves them into code.
Fill out the food profile
A short survey covers allergens, dietary restrictions, and eating style.
Ask follow-up questions
A short conversation tailored to each eating style. Adventurous eaters get a question or two; children with strict food rules get more, so families can start using the tool quickly.
Review and save
Families review and edit the profile before anything is saved.
Find recipes
A family’s search returns the top three recipe matches.
Screen and tailor
Checks for allergens and exclusions, finds substitutions, adapts the recipe, and explains why it works. Recipes with no safe substitution are dropped.
Rate the meal
Families report who ate it, tried it, or refused it, and why.
Suggest profile updates
Turns meal feedback into suggested profile notes.
Set up once
-
01Family
Fill out the food profile
A short survey covers allergens, dietary restrictions, and eating style.
-
02AI
Ask follow-up questions
A short conversation tailored to each eating style. Adventurous eaters get a question or two; children with strict food rules get more, so families can start using the tool quickly.
-
03Family
Review and save
Families review and edit the profile before anything is saved.
Saves
Every meal
Explored, not tested-
04System
Find recipes
A family’s search returns the top three recipe matches.
-
05AI
Screen and tailor
Checks for allergens and exclusions, finds substitutions, adapts the recipe, and explains why it works. Recipes with no safe substitution are dropped.
-
06Family
Rate the meal
Families report who ate it, tried it, or refused it, and why.
-
07AI
Suggest profile updates
Turns meal feedback into suggested profile notes.
Updates go through the same review, looping back to step 03.
07 Evaluation Findings
Testing showed where AI needed a boundary
AI was good at interpreting subjective details like texture, preparation, and similarity. Testing showed it needed structure around it: a defined starting point, and safety rules that don’t depend on the model.
Finding 1
Workflow design mattered more than prompt quality
- What happened
- I first had the AI determine a child’s eating style from open-ended conversation. It wasn’t consistent in how it asked questions or classified the child, and a wrong first call shaped everything after.
- What changed
- Eating style became a survey question. The AI now follows a set line of questioning for each style.
Finding 2
A prompt instruction isn’t a safety guarantee
- What happened
- In the exploratory recipe build, the allergen check lived in the prompt. The model followed it on one run and ignored it on another, producing an unsafe parallel meal.
- What’s changing
- The next build moves the allergen check into code and adds a verification step before any recipe is shown.
08 Where the Work Landed
A working prototype, and a rebuild underway
-
Done
Working prototype
Profile creation and the follow-up chat, tested across 40 cases for four eating styles. Recipe adaptation explored.
-
Recipes from any source
The first build used the Spoonacular recipe API, but it missed family staples like mac and cheese, and many recipe sites block scraping. So families can add recipes their own way:
- Paste a recipe or type your own
- Upload a screenshot or PDF
- Generate with AI
Allergen checks also move into code, verified before any recipe is shown.
-
Later
Test with families
Put the new workflow in front of families and test whether it supports reliable rotation without adding effort. Longer term, the same rules could help anyone at the table, like soft foods for a grandparent with dentures.
09 Reflection
What this taught me about building with AI
Sensory Sprout started as a personalization problem and became a trust problem. Two lessons carry into any AI product: when a wrong answer is costly, people need to see the reasoning and approve what the system learns. And workflow design shapes AI behavior more than prompt wording does.