NutriGym

Overview

NutriGym is a mobile-first, cross-platform fitness ecosystem — built with React Native to run natively on both iOS and Android from a single codebase — that brings workout tracking, nutrition tracking, and body measurements together in one platform. What sets it apart is the LangChain + Google Gemini powered conversational AI trainer and dietician agents, which use the user's past workout and nutrition data as context — giving users personalized, real-time advice that feels like working with a real coach. It's a two-person project, built together with a university colleague and coordinated through Git. The backend was containerized with Docker and deployed to Azure Container Apps as part of our cloud infrastructure. The web admin panel gym staff use to manage members, plans and content shares that same backend, built separately with React, TypeScript and Tailwind CSS and deployed on Vercel.

Tech Stack

API Flow

1. 1The client app sends the request to the Express.js web server.
2. 2The middleware layer runs in sequence: Rate Limiter → Blacklist Check → Auth Guard (JWT validation).
3. 3The API Router dispatches the request to the relevant module's controller.
4. 4The controller delegates business logic to the Service layer; cross-module communication goes through a neighboring module's public service rather than its database directly (low coupling).
5. 5The service reads/writes MongoDB via its Mongoose model and returns the result to the client as JSON.

RAG Flow — AI Coaching Engine

The AI agent's decision logic is modeled as a LangGraph state machine. A user message passes through the following nodes in order:

1. 1Guardrail: The incoming user message first passes a safety check; unsafe content is routed straight to the Response Node and blocked.
2. 2Classifier Node: Safe messages are classified by topic: training, nutrition, or a general question.
3. 3Trainer / Dietician Agent: Training or nutrition queries are routed to the matching agent; if it needs database context it calls a Tool Node (e.g. fetching the user's workout plan), takes the result back through a feedback loop, and generates the reply.
4. 4Retrieval Node + General LLM (RAG): General questions go through the Retrieval Node, which fetches relevant articles from an external knowledge base and feeds them as context to the General LLM — a classic Retrieval-Augmented Generation pattern.
5. 5Response Node: The generated reply is streamed to the client over SSE (Server-Sent Events); a typing indicator and skeleton loader hide the expected latency in the UI.

Performance

Standard API endpoints respond in under 50 ms on average thanks to MongoDB indexes. The AI chat endpoint is far slower due to network latency to Gemini, model inference time, and LangChain classification/routing loops — to hide this from the user, the UI streams the response over SSE alongside a skeleton loader and typing indicator.

Testing & Results

• Three testing levels: Unit (controller/model validations with mocked/stubbed DB and SMTP calls), Integration (cross-module JWT/service communication, LangChain/Gemini routing), and System (use-case-based, end-to-end black-box scenarios).
• Mock dataset: 10 member accounts, 50 exercises, 100 foods/recipes, and predefined trainer availability tables.
• With a 15-user test group, the System Usability Scale (SUS) score reached 88 on the Nutrition screen.
• Task Completion Time dropped across the Home, Nutrition, Profile, and Admin flows — 25→8s, 19→6s, 15→5s, and 18→5s respectively — with the Admin member-registration flow reaching a 100% Task Success Rate.