Vitra

Inspiration

Tracking food intake sucks. The inspiration for Vitra was born from this simple, frustrating truth. While 90% of people start nutrition tracking with good intentions, most quit within three weeks because manual entry is a cognitive tax no one wants to pay. I saw a massive gap between legacy apps like MyFitnessPal (manual and clunky) and newer AI apps like CalAI (vision-only but context-blind). Therefore, I built a Ghost in the Machine, a proactive auditor that doesn't just record what you ate, but understands why it matters to your specific biology, whether it's managing chronic iron deficiencies, or meeting your long-term goals.

What it does

Vitra is the world’s first conversational nutrition coach that utilizes Ambient Intelligence* to eliminate input friction. There's a few components that make it functioning:

Multimodal Logging: Users can snap a photo, type into a chatbot, record a quick voice note (e.g., "I ate half a McDonalds Quarter Pounder"), or do a mix of them all to provide context that barcodes alone cannot see. It tracks both macro and micronutrients.

Health Memory: The agent maintains a persistent understanding of the user's medical profile, such as allergies and micronutrient deficiencies.

Restaurant Grounding: If you're at a restaurant, Vitra uses Gemini 3 Google Search to find the specific menu and calculate precise nutrient profiles in real-time.

Proactive Auditing: Utilizing Gemini 3, it identifies gaps in your day and suggests delivery options to hit your remaining targets, acting as a nutritionist in your pocket.

How I built it

Vitra is built on a high-performance, AI-native stack:

Frontend & Framework: Next.js 14+ with Tailwind CSS and Shadcn UI for a responsive, dashboard-driven experience.

The Brain: Gemini 3 Flash via the google-genai SDK. I utilized Native Multimodality to process images and audio simultaneously and Function Calling for Google Search grounding.

Backend & Auth: Supabase (PostgreSQL) handles user profiles, long-term health context, and meal logs.

Nutrient Logic: I implemented dynamic calculations for TDEE (Total Daily Energy Expenditure) using the Mifflin-St Jeor Equation.

Challenges I ran into

Hallucination: Integrating Google Search Retrieval with a high enough dynamic threshold to find obscure local restaurant menus required significant fine-tuning to ensure Vitra didn't "hallucinate" calorie counts for non-existent dishes.

Ambiguity: A separate challenge was handling the "ambient" nature of human speech. Users don't say, "I had 45 grams of almonds"; they say, "I grabbed a small handful of nuts while running to a meeting." During early testing, the model would either default to the largest possible portion size or ask too many follow-up questions, which broke the "Zero-Friction" promise. To remedy this, I built a Probabilistic Portion Engine within the Gemini 3 logic. I trained the prompt to map casual terms like "a bite," or "a ton of" to specific standard deviation ranges. Instead of forcing a single number, Vitra now logs the most likely mean value but adds a "Contextual Nudge" in the chat, asking, "That was a heavy handful, right?" only if the estimated calorie delta exceeds 200. This kept the UI clean while maintaining scientific integrity.

Accomplishments that I'm proud of

I'm incredibly proud of reducing the "Time to Log" from the industry average of 5 min/day down to just 1 minute for my personal use. Successfully merging a user's long-term medical history (like an iron deficiency) with real-time restaurant search to provide a meal recommendation is a level of "Personalized AI" that doesn't exist in commercial products. Finally, AI-powered suggestions to ensure users meet macro/micro nutrient needs creates high impact that improves the lives of our users.

What I learned

The most important thing I learned was that user retention in health tech isn't about better charts or advanced features; it's about removing the burden of data entry. If the AI can "audit" a meal through ambient conversation, the user stays engaged.

What's next for Vitra

The roadmap for Vitra involves moving from a "reactive" logger to a "real-time" companion:

Gemini Live API: Integrating real-time audio so Vitra can talk back to you while you're grocery shopping or cooking. Wearable Integration: Automatically pulling in active calorie burn from Apple Watch/HealthKit to adjust nutrient targets dynamically. QR Scanning: Implementing high-resolution OCR to audit packaged food labels instantly via the Gemini vision pipeline.

Built With

• gemini
• next.js
• postgresql
• react
• rest
• supabase
• tailwind
• typescript