Beyond Heavy API Keys: How I Built a Zero-Cost Asynchronous AI Summarizer in Spring Boot 3.4

A developer built EchoEngine, a Spring Boot 3.4 application that summarizes YouTube videos using Google's Gemini 2.5 Flash model without API keys for metadata extraction. The app leverages Java 21 virtual threads for high concurrency and uses public endpoints like oEmbed and HTML regex to fetch video metadata. The developer also documented a GCP quota trap where corporate-managed projects enforce zero-quota defaults, requiring migration to an unmanaged sandbox.

When building cloud services around Generative AI, developers often face two major challenges: To address this, I built EchoEngine —a lightweight, high-concurrency Spring Boot 3.4 application designed to ingest YouTube URLs, scrape publicly available metadata cleanly without YouTube API keys, and generate 3-bullet-point technical summaries using Google’s Gemini 2.5 Flash model. In this article, I will break down the system architecture, how Java 21 Virtual Threads Project Loom kept the application non-blocking, and the real-world GCP API quota traps I ran into while building it. spring-ai-starter-model-google-genai <dependency <groupId org.springframework.ai</groupId <artifactId spring-ai-starter-model-google-genai</artifactId </dependency <dependency <groupId org.springframework.boot</groupId <artifactId spring-boot-starter-web</artifactId </dependency 🏗️ Architectural Overview & Core Workflow EchoEngine isolates data fetching, prompt construction, and LLM communication into clear service boundaries. HTTP Client │ ▼ POST /api/summarize EchoController ── Extracts Video ID ──► YouTubeSummaryService │ ┌─────────────────┴─────────────────┐ ▼ ▼ YouTubeClient Spring AI ChatClient oEmbed + HTML Regex Google Gemini 2.5 1. Key-Less YouTube Metadata Extraction Instead of registering for developer API keys, YouTubeClient uses public endpoints: Title: Fetched via YouTube’s public oEmbed endpoint https://www.youtube.com/oembed?url=...&format=json . Description: Extracted via Spring’s RestClient by parsing raw HTML meta tags using regular expressions. 2. High-Concurrency via Java 21 Virtual Threads Because calls to YouTube and Gemini involve substantial network latency, standard OS thread pools risk exhaustion under load. By configuring Virtual Threads in application.properties, every incoming request runs on a lightweight virtual thread: Enable Project Loom Virtual Threads spring.threads.virtual.enabled=true When RestClient blocks waiting for Gemini to finish generating text, the underlying carrier thread is unmounted to perform other tasks, giving the application non-blocking scalability without reactive stream complexity. 3. Spring AI Model Orchestration Spring AI auto-configures a ChatClient bean using settings declared in application.properties: spring.ai.google.genai.api-key=${GEMINI API KEY} spring.ai.google.genai.chat.options.model=gemini-2.5-flash The service constructs a targeted prompt and hands it off cleanly to the LLM: @Service public class YouTubeSummaryService { private final ChatClient chatClient; public YouTubeSummaryService ChatClient.Builder chatClientBuilder { this.chatClient = chatClientBuilder .defaultSystem "You are a technical assistant. Given a video's title and description, " + "provide a 3-bullet-point technical summary." .build ; } public String summarizeVideo String videoUrl { // Fetch metadata via YouTubeClient & pass to Gemini return chatClient.prompt .user promptText .call .content ; } } 🛑 Real-World Debugging & Battle Scars The "Corporate Sandbox" 429 API Quota Trap During early testing, the initial API key continuously threw a 429 Quota Exceeded limit: 0 error despite zero previous requests. The Cause: The key was generated inside a corporate-managed Google Cloud Platform project domain. GCP automatically enforces an explicit zero-quota default limit: 0 on enterprise free tiers until billing accounts are linked. The Fix: Migrating to an unmanaged developer sandbox project in Google AI Studio unlocked standard developer free-tier allowances. Model Override Issues Spring AI's default configuration targeted gemini-2.0-flash out-of-the-box. Since the AI Studio account allocation was explicitly set for the gemini-2.5-flash model endpoint, runtime parameters were passed to force the override: ./mvnw spring-boot:run "-Dspring-boot.run.arguments=--spring.ai.google.genai.api-key=YOUR KEY --spring.ai.google.genai.chat.options.model=gemini-2.5-flash" ⚡ How to Run Locally Set the API Key: $env:GEMINI API KEY="YOUR GEMINI API KEY" Start the Service: ./mvnw spring-boot:run Trigger the POST Endpoint: $body = @{ url = " https://www.youtube.com/watch?v=dQw4w9WgXcQ https://www.youtube.com/watch?v=dQw4w9WgXcQ " } | ConvertTo-Json Invoke-RestMethod -Uri " http://localhost:8080/api/summarize http://localhost:8080/api/summarize " -Method Post -Body $body -ContentType "application/json" 🎯 Key Takeaways Decoupled Architecture: Separating scraping, prompt logic, and controllers makes the codebase maintainable and testable. Virtual Threads simplify high-I/O applications: Project Loom delivers high throughput without rewriting code to be reactive.