Smart library system

A full-stack library management system where users browse, search, and borrow books. Librarians manage inventory and approve requests through an admin panel. AI features natural language search, a streaming chatbot, and post-checkout recommendations.

Live Demo View Source

The Problem

Traditional library systems have two pain points: clunky keyword search that forces users to know exactly what they're looking for, and no way to surface "you might also like" suggestions at the moment a user is most engaged — right after checkout. The goal was to solve both without requiring a backend AI subscription.

The Solution

Built a Next.js App Router application where AI is opt-in: users paste their OpenAI key once into the navbar, it lives only in sessionStorage, and every AI request forwards it via an X-API-Key header. Three AI features were implemented: - Natural language search — sends a compact catalog (~400 tokens) to gpt-4o-mini and gets back ranked book IDs - Streaming chatbot — library assistant with full catalog context, using Vercel AI SDK's useChat + streaming - Post-checkout recommendations — hybrid approach: DB query for genre/author matches, then LLM ranking if a key is available, score-based fallback if not

My Role

Full-stack developer

Tech Stack

Next.js 16 (App Router) TypeScriptReact 19 PostgreSQLPrisma 7Auth.js v5Tailwind CSS v4OpenAIVercel

Key Decisions

1
AI key in the browser, not the server — storing the key in sessionStorage and forwarding it per-request means zero AI infrastructure cost and zero key-leakage risk on the server. The tradeoff is that AI features require the user to bring their own key.
2
Decrement stock on APPROVE, not PENDING — reducing copiesAvailable only when a librarian approves prevents false "out of stock" states for books that might never get approved. All stock changes happen inside Prisma transactions to prevent race conditions.
3
Auth.js Edge split — Next.js middleware runs on the Edge Runtime, which can't use Node.js APIs. Splitting the config into auth.config.ts (edge-safe, no DB) and auth.ts (full Node.js) was necessary to protect routes without breaking the build.
4
Score-based AI fallback — every AI feature degrades gracefully: recommendations fall back to a genre/author scoring algorithm, search falls back to standard text search. The app is fully usable without any API key.
5
Hybrid DB + LLM recommendations — rather than sending the entire catalog to the LLM, a DB query first narrows candidates to in-stock books in matching genres/authors (max 20). The LLM only ranks those 20, keeping prompts small and latency low.

Screenshot Gallery