Vera

Traditional nutrition apps fail because they're rigid. They give users a meal plan, the user breaks it on day three ("I went out for pizza"), and the whole structure collapses into guilt and abandonment. The market is full of calorie trackers that punish deviation and meal-plan apps that assume robotic compliance. The real need: software that absorbs the chaos of real life - a late lunch, a birthday cake, a skipped breakfast - and intelligently recalibrates the rest of the day around it.

A multi-agent LangGraph system coordinated through a state machine, with specialized agents for planning, safety, and conversation. A multi-step tool chain turns a vague request into precise grams: identify the open meal slot, generate a recipe, fit ingredients to the macro target, validate, propose the meal to the user. Hono.js backend, Prisma + Supabase Postgres, native SwiftUI iOS app with SSE-streamed chat. Macro data sourced from public nutrition databases.

Live on the App Store. The agent pipeline runs against real macro data in production, and a ring-based reinforcement layer - the ring is the reward - now closes the loop on top of the original meal-fitting agent: USDA-enriched micros, a daily focus-ring card, and grounded coaching projections.