Devil Fruit Career Advisor

Provides personalized career recommendations by matching user personality traits to "Devil Fruit Powers" and subsequently to "Enterprise Job Roles."

Comment

Inspiration

As a devoted One Piece fan for nearly 10 years, I've always been fascinated by the central question that every fan asks themselves: "What Devil Fruit would suit me?" This quirky anime concept became the spark for a much deeper exploration.

But I wanted to create something beyond just a fun fandom project. As a student actively searching for job opportunities, I deeply understand the anxiety and uncertainty of career exploration. The question "What role is right for me?" resonates just as strongly as "What Devil Fruit fits my personality?"

That's when the vision clicked: What if I could blend the imaginative world of One Piece with real-world career guidance? By mapping Devil Fruit abilities to professional traits and job roles, I could build an AI-powered career advisor that's both engaging and genuinely impactful—helping users discover career paths that align with their unique strengths, just like finding the perfect Devil Fruit.

How we built it

How we built it The project implements a 4-step agentic workflow powered by NVIDIA NIMs. Step 1 uses custom guardrails logic to validate user input for safety and relevance. Step 2 leverages NV-Embed-QA to generate 2048-dimensional semantic embeddings of user narratives, performing RAG retrieval against 230 Devil Fruits and 230 job roles stored in AWS S3. Step 3 executes a hybrid tool strategy—calling an AWS Lambda function for static market data while simultaneously enriching results with Google Custom Search insights (both depth and breadth queries). Step 4 feeds all context into Llama 3.1 Nemotron 70B Instruct for reasoning, then applies a deterministic reflection engine that scores recommendations across trait alignment, difficulty match, tool data usage, and justification quality. The frontend is a vanilla JavaScript SPA hosted on S3, with API Gateway endpoints for secure credential retrieval via AWS Secrets Manager. Performance tracking and agent trace visualization provide full observability into each workflow step.

Challenges we ran into

The biggest challenge was discovering I lacked AWS Marketplace subscribe permissions hours before submission. My original plan to deploy NIMs via SageMaker endpoints became impossible, forcing an immediate pivot to NVIDIA's hosted API key approach. This meant rewriting the entire inference layer under extreme time pressure. Additionally, since Nemo Guardrails isn't available as a public API, I had to implement custom guardrails validation from scratch. Another hurdle was RAG retrieval performance—my S3 data lacked pre-computed embeddings, so computing vectors for 460 items on-the-fly would take 5-10 minutes per run. I solved this with a hybrid approach: generating embeddings only for user input (demonstrating NIM usage) while falling back to keyword matching for item comparison. The hackathon manager eventually granted permissions, but too late to reverse my architectural decisions—teaching me that shipping a working prototype beat chasing the perfect solution.

Accomplishments that we're proud of

I'm incredibly proud of building a production-ready agentic AI system in such a short timeframe while learning AWS from near-scratch. The 4-step workflow successfully demonstrates advanced AI patterns: semantic search with Embedding NIM, hybrid tool orchestration (static + dynamic web data), and two-phase reasoning with deterministic validation. Despite the permission roadblocks, I delivered a fully functional prototype that generates meaningful career recommendations with 0.75+ verification scores and ~15-second end-to-end latency. The system gracefully handles errors, implements secure credential management via Secrets Manager, and provides rich observability through performance metrics and agent trace timelines. Most importantly, I merged my passion for One Piece with genuine career impact—the project isn't just a novelty; it's a tool I'd actually use as a job-seeking student. Successfully pivoting from SageMaker to API keys under pressure proved I can adapt and deliver even when plans fall apart.

What we learned

This hackathon taught me invaluable lessons across three dimensions. Technically, I mastered multiple AWS services I'd never touched before—Lambda functions, S3 static hosting, API Gateway, Secrets Manager, and CloudWatch—while learning to integrate NVIDIA NIMs for production AI workflows. I gained deep understanding of RAG architectures, semantic embeddings, and the nuances of prompt engineering for structured JSON outputs. Architecturally, I learned that graceful degradation and fallback strategies are just as important as the happy path; my hybrid retrieval approach (embeddings + keywords) kept the system functional despite missing pre-computed vectors. Strategically, the most critical lesson was knowing when to pivot—spending hours debugging SageMaker permissions would have meant no submission at all, while switching to API keys took 30 minutes and unblocked everything. Time-boxed decision-making under pressure is a skill I'll carry into every future project.

What's next for Devil Fruit Career Advisor

The immediate priority is pre-computing embeddings for all 460 items and storing them in a vector database (Pinecone or Weaviate) to enable true semantic similarity matching instead of keyword fallback. I'll deploy real Nemo Guardrails as a containerized microservice for production-grade safety filtering, including jailbreak detection and prompt injection prevention. Once I have proper AWS permissions, I'll migrate to SageMaker endpoints for Llama 3.1 and Embedding NIM to reduce latency and eliminate external API dependencies. For user experience, I'll add authentication via Cognito to enable personalized career history tracking and multi-year progression roadmaps. The system will integrate real-time job market data from LinkedIn and Indeed APIs, updating demand scores daily and alerting users to new relevant postings. Finally, I'll implement multi-modal input (resume upload, LinkedIn profile import) and build a mobile-responsive PWA with offline support. Long-term, I envision this evolving into a comprehensive career companion that grows with users throughout their professional journey—blending the fun of Devil Fruit discovery with actionable career guidance.

Share this project:

Updates