Fridge Chef

Inspiration

Every evening, the same conversation happens in millions of households:

"What are we cooking tonight?" "I don't know, what do we have?" "...let's just order takeout."

I've been that person more times than I'd like to admit. The frustration isn't that I can't cook — it's that the gap between "what's in my fridge" and "what I should make" is exhausting to bridge. Recipe apps assume you'll go shopping for their ingredients. Cooking videos assume you have everything already. Nothing meets you where you actually are: standing in front of an open fridge, tired, at 7 PM.

And there's a second frustration nobody seems to solve: when I do plan multiple dishes for the week, I end up with three half-overlapping shopping lists, buying the same scallions twice and forgetting the soy sauce entirely.

Fridge Chef is the app I wish existed. One photo of your fridge → tonight's dinner, this week's meal plan, and one consolidated shopping list — all from a conversation.

What it does

Snap your fridge — Gemini's vision model identifies every ingredient as editable tags.
Get 6 dish suggestions — Each one shows what you already have vs. what you'd need to buy.
Multi-select recipes — Pick the dishes you want to make this week.
One auto-merged shopping list — Missing ingredients are deduplicated, grouped by category, with notes on which dishes need them.
AI-illustrated steps — Every key cooking step gets a photorealistic image generated by Nano Banana.
Conversational tweaks — "No cilantro." "Under 15 minutes." "Make it vegetarian." The recipe rewrites itself.

How I built it

The whole app was built without writing a single line of code, through multi-turn conversations with MeDo.

I structured my MeDo conversations in three rounds, which turned out to be far more reliable than trying to specify everything upfront:

Round 1 — Skeleton. I described the four core pages (upload, ingredient confirmation, recipe grid, recipe detail), the data model, the visual tone ("warm kitchen, not cold tech"), and explicitly told MeDo what not to build (no auth, no comments, no social features). MeDo generated the full-stack scaffold — frontend, backend routes, and database tables — in one shot.

Round 2 — AI integrations. I added the model pipeline:

Gemini Vision → analyzes the fridge photo and returns a structured list of ingredients
LLM-driven recipe generation → takes the ingredient list, returns six dishes as structured JSON (including English dish names and short visual descriptions for the image step)
Nano Banana (Gemini 2.5 Flash Image) → generates the hero shot and step-by-step images, with the hero shot passed as a style reference to maintain visual consistency

The single biggest performance unlock was telling MeDo to batch all step images in parallel instead of generating them sequentially. The difference between 30 seconds of waiting and 4 seconds is the difference between a usable product and a tech demo.

Round 3 — The killer feature. I added multi-select on the recipe grid, a sticky cart-style bottom bar, and an aggregated shopping list page that deduplicates ingredients across recipes, groups them by category (vegetables / meat / seasonings / staples), and annotates each item with the dishes it's used in.

Challenges I ran into

Ingredient recognition hallucinated. Early on, Gemini would confidently identify "kale" in a fridge that contained zero kale. Solution: never trust, always show. Every recognized ingredient is an editable tag — one tap to remove, one tap to add a missed item. The UX turned a model limitation into a feature.

Step images looked like four different shoots. When I generated each step image independently, every photo had a slightly different lighting, color palette, and plate style — the recipe looked like a Pinterest collage. The fix was to generate the hero shot first, then pass it as a reference image to every subsequent step generation, with the prompt instruction "maintain the same lighting, color palette, and overall photographic style as the reference image." The whole recipe now looks like a single coherent photo essay.

Multi-turn refinement kept losing context. When users said "now make it vegetarian," the model would forget the previous dish entirely and start fresh. Fix: pass the full current recipe state + the user's edit request as a single coherent prompt, not just the latest user message.

Shopping list deduplication is harder than it looks. "Scallions" and "green onions" and "葱" are the same thing. String matching fails. Solution: use the LLM itself as a semantic deduplicator — it merges synonyms and dialectal variants into a canonical list.

Accomplishments I'm proud of

A production-deployed full-stack app in 9 days, zero code written by hand.
The multi-select → aggregated shopping list flow, which I haven't seen any recipe app actually solve well.
Visual cohesion across AI-generated images that genuinely looks like one shoot.
Parallel image generation that makes the experience feel instant rather than waiting on AI.

What I learned

Prompting an AI no-code platform is its own skill. The mental model that worked best for me was treating MeDo like a brilliant junior engineer: be specific, give constraints (especially what not to do), iterate in small rounds, and review output rather than trying to one-shot the perfect spec.

Constraints are a feature, not a limitation. Telling MeDo "no auth, no social, no nutrition calculations, mobile-first only" actually accelerated everything. The narrower the scope, the better the output.

Multi-model pipelines are now trivial. A year ago, wiring Gemini Vision → LLM recipe generation → Nano Banana image generation → database persistence would have been a weekend of glue code. With MeDo, it was a paragraph of plain English.

What's next for Fridge Chef

Memory of dietary preferences — allergies, dislikes, household size remembered across sessions.
Weekly meal planner — extend the multi-select pattern from "tonight" to a full Mon–Sun plan.
One-tap grocery delivery integration — turn the shopping list into a Meituan / JD Daojia order.
Family-shared lists — collaborate on what to cook with whoever else is in your household.