How AI runs NutriSight
Every prediction, recommendation and lesson on NutriSight is generated by an AI agent running in production today. This page lists those agents, what they decide on their own, and where the human founder is still in the loop.
Live counters
What the system has done so far
—
Users
—
AI scans analyzed
—
Gemini calls logged
—
Lessons completed
—
Beta signups
Pulled live from the production database. Numbers are real, not simulated.
Product agents
AI that runs the product
| Agent | Status | Decides autonomously | Human still owns |
|---|---|---|---|
Food Recognition Agent | Live | Identifies dishes, ingredients and portions from camera images using Google Gemini 2.5 Flash multimodal vision. | Founder reviews edge-case misclassifications weekly to refine prompts. |
Menu Understanding Agent | Live | Reads café and restaurant menus from photos, picks the best dish for the user's goal. | Founder curates representative menu test set for QA. |
Nutrition Estimation Agent | Live | Estimates calories, sugar, carbs, protein, fiber and salt per serving via Gemini reasoning. | Founder spot-checks against USDA / Frida database for calibration. |
Recommendation Agent | Live | Generates a personalized recommendation based on the user's goal and the meal scanned. | Founder writes goal taxonomy and safety guard prompts. |
Meal Memory Agent | Live | Stores meals, computes daily totals, decides the next-best-action card on Today. | Founder owns the daily target heuristics until learned per-user. |
Coach & Learning Agent | Live | Surfaces lessons matched to the user's current goal and recent scans. | Founder approves new lesson content before publish. |
Safety Agent | Live | Adds disclaimers, blocks medical-claim phrasing, refuses unsafe requests. | Founder reviews flagged outputs weekly. |
Business agents
AI that runs the company
| Agent | Status | Decides autonomously | Human still owns |
|---|---|---|---|
Support Agent | Beta | Drafts first-response replies to user emails and in-app feedback using Gemini, suggests resolutions. | Founder reviews and sends every reply during beta. |
Onboarding Agent | Beta | Personalizes a user's goal & starting lessons based on their signup answers. | Founder owns the welcome email tone & cadence. |
Content Agent | Planned | Will draft new Academy lessons weekly from a research outline and validate citations. | Founder reviews and publishes — no auto-publish. |
Weekly Report Agent | Planned | Will generate each Pro user's weekly nutrition report from their saved meals. | Founder reviews the prompt and sample reports before rollout. |
We label every agent honestly as Live, Beta or Planned. Nothing on this page is aspirational marketing.
Stack
The Google Cloud stack behind NutriSight
Google Gemini 2.5 Flash & 2.5 Pro — multimodal vision + reasoning for every food scan, menu read, recipe parse, comparison and meal description.
Lovable AI Gateway routes every Gemini call so we can log model, latency and outcome per request (visible to the founder in the Admin dashboard).
Lovable Cloud (managed Postgres + Auth + Storage) stores meal memory, learning progress, and scan events with row-level security per user.