Inspiration
Restaurant ordering is one of the hardest moments for people trying to stay consistent with nutrition goals. Calorie trackers are useful after a meal, but they usually do not help when you are staring at a menu, missing protein data, comparing prices, and trying to make a quick decision before you order.
MacroTune moves nutrition assistance upstream. Instead of logging regret after the fact, it helps users make a smarter restaurant order while the decision is still changeable.
What it does
MacroTune is an AI-powered nutrition assistant that turns restaurant menus into optimized meal recommendations.
A user sets three practical constraints:
- Daily calorie target
- Daily protein target
- Meal budget
Then they import a menu by photo, pasted text, or URL. GPT-5.6 parses the menu into structured items and estimates missing nutrition where needed. Estimated values are clearly labeled so the user knows what is official, inferred, or editable.
After review, MacroTune runs a deterministic optimizer that evaluates bounded meal combinations of up to three menu items. It returns the top three recommendations that maximize protein while staying within calorie and budget constraints. The user can log the selected meal, and the dashboard updates remaining calories, remaining protein, and seven-day history locally.
How we built it
MacroTune is designed as a mobile-first Next.js and TypeScript single-page PWA. The MVP intentionally uses localStorage only, with no backend, database, or authentication. That keeps the product fast, private, and easy to demo.
The AI boundary is explicit:
- GPT-5.6 parses messy menu input into structured JSON.
- GPT-5.6 estimates missing nutrition values and labels uncertainty.
- GPT-5.6 writes concise tradeoff tips after recommendations are ranked.
- Deterministic TypeScript, not the language model, chooses the meal recommendations.
This separation makes the product more trustworthy: AI handles messy interpretation, while repeatable code handles the numeric optimization.
How Codex helped
Codex was used as a senior product, design, and engineering collaborator throughout the project. It helped shape the product strategy, define the PRD, clarify the AI architecture, preserve the deterministic optimizer boundary, and prepare implementation-ready documentation for the app experience.
Codex accelerated the workflow by turning a raw product idea into a structured product system: target users, jobs to be done, journey maps, recommendation-screen requirements, accessibility guidance, technical boundaries, and hackathon-ready positioning.
Key product decisions
- The recommendation screen is the primary demo moment.
- The app optimizes before ordering, rather than only tracking after eating.
- Missing nutrition data is estimated but always labeled.
- Users can manually edit extracted menu data so parsing failures never block the workflow.
- Recommendations are deterministic and constrained by calories, protein, and budget.
- The MVP avoids accounts, backend storage, restaurant integrations, payments, and social features.
What makes it different
Most nutrition apps are food diaries. Most delivery apps optimize for ordering conversion. MacroTune is a decision layer between those worlds: it reads a real menu, understands the user's constraints, and recommends practical meal combinations before money is spent and calories are consumed.
The result is a calmer, faster way to order takeout while staying aligned with macro goals.
Challenges we ran into
The core challenge was designing an AI workflow that remains trustworthy when restaurant nutrition information is incomplete. The solution was to avoid letting the language model make final nutrition decisions. MacroTune uses GPT-5.6 for extraction and explanation, then hands structured data to deterministic optimization logic.
Another challenge was scope control. It would be easy to add accounts, food databases, delivery integrations, full meal planning, or wearable sync too early. The MVP stays focused on one repeated high-friction job: given this restaurant menu, this calorie target, this protein target, and this budget, what should I order?
Accomplishments that we're proud of
MacroTune has a clear product wedge, a strong user problem, and a credible AI architecture. It treats AI uncertainty honestly instead of hiding it, keeps the user in control through editable menu data, and makes the recommendation logic explainable.
The project also has startup-quality product documentation behind it, including a detailed PRD that can guide design, engineering, QA, demo preparation, and future roadmap decisions after the hackathon.
What's next
Future versions can add saved restaurants, remembered user edits, dietary preferences, tracker exports, Apple Health summaries, official restaurant nutrition ingestion, and eventually commerce-adjacent ordering partnerships.
The long-term vision is for MacroTune to become the intelligent nutrition layer between users and every food purchasing decision: restaurants first, then grocery, cafeterias, travel meals, and household food planning.
Built With
- codex
- deterministic-optimization
- gpt-5.6
- localstorage
- next.js
- openai-api
- pwa
- typescript