SkillMiner

Inspiration

This year, we’ve seen an explosion of new AI agents—yet most people still use them passively, asking questions instead of letting agents actively help them do things. We wanted to change that. Our vision is to make agents feel like independent digital assistants who can take initiative, execute actions, and continuously support users in achieving their goals rather than just responding to prompts.

As students, we realized one of the hardest challenges is preparing for jobs efficiently. While existing job platforms focus on optimizing résumés to pass ATS filters, they rarely help students learn the right skills or bridge the gap between where they are and where they want to be. Students often waste hours researching multiple resources and manually designing study plans.

SkillMiner aims to solve this by building an AI agent and learning platform that can understand your career goals, analyze your résumé, and automatically generate a personalized learning roadmap—complete with milestones, deadlines, and study materials that align with your schedule and job target.

What it does

SkillMiner is an AI-powered platform that combines résumé analysis, skill-gap detection, and adaptive learning planning into one seamless experience.

After logging in with Google, users upload their résumé. The system:

Analyzes résumé content to extract technical and soft skills.
Compares these skills with real job descriptions (e.g., “Data Analyst,” “Machine Learning Engineer”) to find missing capabilities.
Generates a personalized study plan with specific tasks, estimated durations, and categories such as Practice or Study.
Displays these tasks in a daily dashboard—each with XP values and quick-action buttons like “Start Task” or “Reschedule.”
Offers a real-time AI study assistant that chats with you to modify plans, swap learning resources, or explain topics directly inside the interface.

Every interaction strengthens your learning path. The more you engage, the more the system learns about your pace, preferences, and interests—building a continuously evolving plan tailored just for you.

How we built it

Frontend: React + TypeScript + Tailwind CSS (Vite). The interface includes a modular dashboard, XP progress tracker, and a persistent chat agent.
Backend: Supabase is used for authentication, user session management, and real-time synchronization between frontend and backend. We also leverage Supabase Edge Functions for secure serverless API execution, handling résumé parsing, plan generation, and database updates without exposing sensitive keys.
Data Pipeline: All user and skill-related data flows through an ETL pipeline built with Python Polars, performing data cleaning, enrichment, and transformation. The processed results are stored in PostgreSQL hosted on AWS RDS, ensuring durability, scalability, and fast access for downstream services.
Cloud Infrastructure: The system integrates Supabase (for edge computing, auth, and storage) with AWS services (RDS for relational data and S3 for backup storage), creating a hybrid cloud environment that balances performance, cost efficiency, and security.
Vector Database: We use ChromaDB to vectorize résumé text, job descriptions, and skill documents. This enables fast semantic search and context retrieval for the AI agent to generate accurate and personalized recommendations.
AI Engine: The OpenAI API powers résumé parsing, skill mapping, and personalized task generation through prompt chaining, embedding similarity search, and fine-grained reasoning.
Agent System: The in-app assistant runs as a persistent React component that communicates with Supabase Edge Functions and ChromaDB. It supports contextual commands such as reschedule, swap resource, or explain this concept, dynamically updating the user’s study plan in real time.

Together, these components create a complete end-to-end learning intelligence system—from data ingestion to personalized agent interaction.

Challenges we ran into

Pipeline Orchestration: Ensuring smooth data flow from the ETL process into PostgreSQL and ChromaDB without latency required multiple retry and caching layers.
Résumé Normalization: Handling varying résumé formats to extract clean, comparable skill vectors.
Context Management: Maintaining relevant vector embeddings while ensuring the agent doesn’t overload with irrelevant context.
Real-Time Synchronization: Keeping the UI, Supabase session state, and AWS database aligned under high update frequency.

Accomplishments that we're proud of

Designed a full-stack AI platform that integrates structured database systems with real-time vector search for adaptive learning.
Built a robust data pipeline connecting AWS S3 (PostgreSQL), ChromaDB and Supabase, ensuring scalable skill retrieval.
Developed a chat-based agent that actively modifies study plans through user conversation.
Implemented a gamified XP system that visualizes learning progress and increases motivation.
Created a clean, production-ready UI combining modern design and functional depth.

What we learned

Through SkillMiner, we learned how to operationalize AI intelligence through pipelines rather than one-off models. Building the system forced us to think deeply about data architecture, embedding management, and agent memory. We also learned how to make LLMs not only generate text but manage structured decisions—translating user input into database-backed actions like scheduling and difficulty scaling.

Most importantly, we learned that the real power of AI agents lies not in conversation, but in execution—helping users get tangible results.

What's next for SkillMiner

Memory-Enhanced Agent Architecture: We plan to extend our agent’s capabilities by adding short-term and long-term memory modules. This will allow SkillMiner to retain user context over time—remembering study habits, past mistakes, and long-term goals—making its recommendations increasingly adaptive and personalized.
MCP Integration for Calendar Awareness: By integrating the Model Context Protocol (MCP), the agent will gain access to tools such as Google Calendar. This will enable it to understand users’ real schedules and automatically optimize study plans around upcoming events, deadlines, and free time slots.
Multi-LLM Quality Supervision: We aim to introduce a multi-model supervision layer, where different LLMs evaluate each other’s responses. This scoring mechanism will monitor answer quality and consistency, ensuring that the agent continuously improves its reasoning accuracy and output reliability over time.