Gutsy: Eating with AI

Inspiration

Living with IBS or Colitis means every meal is a gamble. We watched friends and family anxiously Google ingredient lists at restaurants, only to still get blindsided by a flare-up hours later. The information exists — FODMAP charts, trigger databases, elimination diet guides — but none of it is personalized or available in the moment you need it most: right before you eat. We wanted to build something that puts a gut health expert in your pocket, one that actually knows your body.

What it does

Gutsy lets you snap a photo of any food — a plate, a menu, or a nutrition label — and instantly get a personalized risk analysis (green/yellow/red) based on your specific diagnosed conditions and known triggers. It identifies ingredients, flags FODMAP groups, and estimates nutritional info. Beyond scanning, Gutsy tracks your meals and symptoms over time, then uses AI to find Hidden Culprits — the foods that triggered a flare-up hours ago that you'd never suspect. A Community Insights feature connects you with your "Gut Twins," showing what works for people with the same profile.

How we built it

• Next.js 16 with React 19 and the App Router for a fast, modern frontend
• Google Gemini 2.5 Flash for real-time food image analysis with structured JSON output and FODMAP categorization
• Google Gemini 2.5 Pro with extended thinking for the deeper Hidden Culprit detection and Community Insights — tasks that benefit from chain-of-thought reasoning
• PWA (Progressive Web App) via next-pwa so users can install it on their phone and use it offline
• Browser local storage for privacy-first data persistence — your gut health data never leaves your device
• Tailwind CSS 4 and Lucide icons for a clean, mobile-first UI
• Recharts for symptom intensity visualization over time

Challenges we ran into

• Structured AI output: Getting Gemini to consistently return valid, typed JSON responses with the exact schema we needed (meal names, FODMAP groups, confidence scores) required careful prompt engineering and response schema enforcement.
• 48-hour correlation: Building the Hidden Culprit feature meant reasoning about digestion transit times (2–24 hours) across interleaved food and symptom logs — a non-trivial temporal reasoning task even for LLMs.
• Image versatility: The scanner needed to handle wildly different inputs — a blurry photo of pad thai, a close-up of a nutrition label, a restaurant menu — and extract meaningful ingredient data from all of them.
• Balancing speed vs. depth: We used Flash for the scanner (speed matters when you're standing in line) and Pro with extended thinking for insights (accuracy matters when diagnosing culprits).

Accomplishments that we're proud of

• A fully functional food scanner that goes from photo to personalized risk assessment in seconds
• The Hidden Culprit engine that correlates symptoms with meals across a 48-hour window — something most gastroenterologists do manually
• Privacy-first architecture where all user data stays on-device in local storage
• A polished, installable PWA that feels native on mobile
• The "Mark as Safe" feature that lets the AI learn your personal safe foods over time

What we learned

• Structured output from multimodal AI is incredibly powerful — Gemini's response schema feature let us build reliable, typed pipelines without brittle regex parsing
• Different reasoning depths for different tasks: Flash for fast scans, Pro with thinking for deep analysis — matching model capability to task complexity saves cost and improves UX
• Gut health is deeply personal: No two IBS patients have the same triggers, which is exactly why generic apps fail and personalized AI succeeds
• PWAs are underrated: With next-pwa, we got installability, offline support, and a native feel without touching app stores

What's next for Gutsy: Eating with AI

• Wearable integration: Connect with smartwatches to correlate stress levels and sleep with flare-ups
• Photo history matching: "You ate this same dish 3 weeks ago and had a flare-up" — visual memory for your gut
• Collaborative profiles: Share anonymized trigger data to build a real community database (replacing the current AI-generated stats)
• Dietitian handoff: Export a structured report of your logs, triggers, and AI insights to share with your healthcare provider
• Multi-language menu scanning: Traveling abroad shouldn't mean guessing what's safe to eat
• Barcode scanning: Instant lookup of packaged foods against your personal trigger profile

Built With

• nextjs
• pwa