Overview
AniRec AI is a full-stack web application that combines natural language AI recommendations with personal anime library management. Built with React, TypeScript, Node.js, and SQLite, it uses Google Gemini Pro for narrative-driven recommendations and Gemini Flash for fast metadata extraction.
🎯 Problem & Motivation
Standard anime search platforms rely on rigid genre tags. Finding recommendations based on subtle narrative themes, character arcs, or specific pacing is difficult. I wanted an AI assistant capable of understanding natural language preferences.
⚙️ Technical Constraints
External anime APIs (AniList, MyAnimeList, Kitsu) enforce strict rate limits and return disparate data formats. Querying large AI models for simple search parsing introduces unnecessary latency.
🏗️ Architecture & Dual-AI Pipeline
- Gemini Flash Engine: Extracts anime titles, genres, and intent from user search prompts at low latency.
- Gemini Pro Recommender: Evaluates user preferences against anime themes to generate personalized recommendations with explanation rationale.
- API Normalizer: Unifies external anime metadata structures into a consistent internal schema.
- SQLite Database: Local structured storage for watch-list persistence.
⚖️ Challenges & Trade-offs
Trade-off: Using two separate LLM models (Flash + Pro) increases backend pipeline coordination, but drastically improves UI responsiveness compared to routing all queries through a single large model.
💡 Lessons Learned
1. Decoupling parsing from recommendation generation optimizes both speed and token cost.
2. Building a central normalization layer upfront prevents UI state bugs when aggregating external APIs.
📌 Planned Improvements
- Planned Vector similarity search for instant offline recommendations.
- Planned Export/Import watch-lists via JSON format.