Back to selected work

Case study · 01 / 2025 · Product & GenAI

An AI-powered Vietnamese recipe finder backed by Google Gemini 2.5-flash for intelligent recipe generation and smart natural language search.

101+
Curated Vietnamese Dishes
100%
JSON Schema Validation
< 1.4s
Gemini 2.5 Flash Latency
AI Culinary Companion
Role
Product & GenAI
Year
2025
Outcome
101+ Vietnamese Recipes
Architecture Stack
React 19TypeScriptTailwindGemini 2.5-flash

01 / Context & Problem Statement

A modern web application designed to help users search, filter, and discover traditional Vietnamese recipes. Traditional recipe finders are often static and hard to navigate, but this project combines a rich dataset of 101+ Vietnamese dishes with an AI culinary assistant powered by Google Gemini 2.5-flash to generate custom, structured recipes on the fly.

02 / System Architecture & Pipeline Flow

STAGE 0101 / 04

Ingredient & Intent Input

Users specify pantry ingredients, dietary constraints, cooking time, or natural language cravings.

STAGE 0202 / 04

Hybrid Search & Prompt Context

Matches local 101+ Vietnamese dish index or constructs a constrained system prompt for Gemini 2.5-flash.

STAGE 0303 / 04

Structured JSON Synthesis

Enforces strict responseSchema via @google/generative-ai SDK so output never breaks UI rendering.

STAGE 0404 / 04

Interactive Step-by-Step UI

Renders dynamic portion scaling, prep timers, nutritional breakdown, and regional culinary notes.

03 / Engineering Approach

Built using React 19, TypeScript, and Vite for a highly responsive, modern frontend. The AI assistant leverages the official @google/generative-ai SDK, configured with structured output formatting to guarantee that the recipes generated by Gemini always return valid JSON matching our application's design system. Used Tailwind CSS for a responsive grid layout, supporting detailed filters such as cooking time, difficulty levels, dietary restrictions, and portion sizes.

04 / Key Engineering Deliverables

  • Deterministic JSON schema enforcement with Google Gemini 2.5-flash SDK
  • Multi-dimensional filtering across region, prep time, difficulty, and diet
  • Zero-runtime-crash fallback handling for malformed LLM token streams

05 / Engineering Reflection

"Integrating LLMs in user-facing applications requires strict output schema validation. Working with Gemini's structured output mode taught me how to constrain LLM responses to prevent UI breakage while delivering personalized recipes in real-time."

Interested in this architecture? · Let's talk

Let's build
something useful.

[email protected]→→