Bezede: Ethiopian Nutrition & Fitness

Overview

Bezede is the first-ever nutrition and fitness tracking application built specifically for Ethiopian food and lifestyles. Before Bezede, no digital tool could accurately track the caloric and macro content of traditional Ethiopian meals — foods like Injera, Shiro, Tibs, and Kitfo had zero reliable nutritional data in any existing health app.

This gap wasn't just a product oversight; it reflected a deeper problem: the nutritional values of Ethiopian foods were never digitized at scale. I personally sourced, cleaned, and structured the Ethiopian Food Composition Table (2025) — a comprehensive government-issued dataset — transforming it from raw, noisy research data into a production-ready, queryable database that powers every AI inference in the app.

A First of Its Kind

No competitor exists. Every mainstream fitness app (MyFitnessPal, Cronometer, Lose It) fails Ethiopian users because their food databases simply don't include traditional dishes. Bezede solves this permanently by:

• Curating the dataset by hand — cleaning ambiguous entries, normalizing portion sizes, and validating macro values against research sources
• Building the first AI pipeline that can identify Ethiopian foods from photos and map them to exact database entries with confidence scores
• Delivering Amharic-language coaching so users can interact with a nutritionist AI in their native tongue

Key Features

AI-Powered Food Logging (Vision + NLP)

Users upload a photo of their plate or describe a meal in natural language. An AI pipeline chains two models:

1. BLIP (Salesforce) — a vision model that generates a raw text description of the image
2. Llama-3 (Meta) — an LLM prompted with Ethiopian portion logic that cross-references the description against the food database, returning a typed JSON object with weight estimates and macro-nutrient data including confidence scores

Multilingual AI Nutritionist

An integrated chat assistant provides culturally-tailored dietary advice, recipe suggestions, and motivational coaching in both English and Amharic, based on the user's logged intake and biometric profile.

Dynamic Nutritional Dashboard

Detailed daily dashboards tracking protein, carbohydrates, fats, and total calories — benchmarked against personalized metabolic targets (BMR/TDEE) calculated from user biometrics.

Comprehensive Ethiopian Food Database

Instant querying against a curated set of hundreds of Ethiopian dishes with detailed macro and micro-nutrient profiles — the only database of its kind available in a consumer application.

Architecture

Bezede uses a modern, decoupled architecture:

• Frontend (React SPA) — handles CRUD operations against the backend and directly orchestrates LLM inference tool calls for low-latency AI responses
• Backend (Hono + Node.js) — lightweight REST API routing requests to the PostgreSQL database with Better Auth for session management
• AI Orchestration Layer — a multi-provider architecture supporting fallback routing between Gemini, HuggingFace, OpenRouter, and local Ollama inference, preventing vendor lock-in

Notable Implementations

Dual-Layer Dietary Mapping Pipeline

The most technically innovative feature: a data-validation strategy that chains BLIP (vision) → Llama-3 (NLP) via custom prompt engineering. This converts ambiguous images of culturally specific meals into cleanly typed JSON macro-nutrient objects — a pipeline that doesn't exist anywhere else for Ethiopian food.

Extensible Multi-Provider AI Architecture

Built a unified chat API interface that routes between Gemini, HuggingFace Inference, OpenRouter, hasabAi, and local Ollama. If one provider is unavailable or rate-limited, the system falls back seamlessly — eliminating single-provider dependency.

Data Engineering & Curation

Migrated from raw government research documents to a structured, validated PostgreSQL schema via Drizzle ORM. This involved resolving duplicates, normalizing measurement units, and filling data gaps through cross-referencing multiple nutritional sources.

Legacy Architecture Migration

Successfully refactored an older NestJS backend into a streamlined Hono API while maintaining Redux state compatibility and uninterrupted JWT session management for existing users.

Tech Stack

• Frontend: React 19, TypeScript, Vite, Redux Toolkit, React Router DOM, Recharts, Tailwind / Custom CSS
• Backend: Node.js, Hono, Better Auth
• Database / ORM: PostgreSQL, Drizzle ORM
• AI & Integrations: HuggingFace Inference API (BLIP, Llama-3), Google GenAI, OpenRouter, hasabAi, Ollama

Use Case

Designed for Ethiopians and diaspora communities pursuing health, weight loss, or muscle gain goals. Bezede bridges a massive real-world gap in health tech by enabling users to track complex traditional meals that no other application can handle — while delivering native Amharic coaching and culturally intelligent dietary guidance.