Overview

Decision fatigue kills the motivation to cook before the pan even heats up. Food Tinder removes choice paralysis by turning meal planning into a swipe — one recipe at a time, right to save, left to skip. Preference is easier to express than to articulate, and the app learns from every swipe.

Recommendation Engine

A lightweight linear model maps each recipe to a feature vector of ingredient clusters and regional cuisine. Every swipe updates the weights: likes nudge up by 0.15, passes nudge down by 0.08, with a full retrain every 50 swipes to correct drift. The asymmetric rate reflects how preferences work — aversion is weaker than attraction.

To avoid echo chambers, picks use an 80 / 20 exploit-explore split: 80 % highest-scoring unseen recipe, 20 % random — surfacing cuisines the model hasn't formed a fair opinion on yet.

Explainability

Every card shows a match percentage derived from the current weight vector, making the model's reasoning legible. A 91 % on Butter Chicken signals strong protein, dairy, and Indian weights; 34 % on Ratatouille means the model learned you deprioritise vegetable-dominant dishes. It shifts the mental model from "the app picked this" to "I trained the app to pick this".

Try It Live

Runs entirely in your browser — the model starts learning immediately.