• 

  AI agent page

• 

  Agent in Action

• 

  Dreamers Agent landing page

• 

  Agent's output format

Inspiration

As an immigrant and first-generation college student, I know firsthand how overwhelming it can be to find scholarships, financial aid, and resources while balancing school and work. I often wished there was a tool that could understand my unique situation and guide me toward opportunities I qualify for. That personal struggle inspired me to create Dreamers Agent, an AI-powered assistant built to simplify the process for students like me. The project is driven by the idea of building with purpose — making technology that doesn’t just work, but truly helps people who need it most.

What We Learned

Working on Dreamers Agent deepened my knowledge in several areas:

• How to integrate large language models (LLMs) with real-world data retrieval.
  
• The importance of prompt engineering and designing structured responses for clarity.
  
• How to balance usability and accuracy in AI applications.
  
• Building for scalability — ensuring the system could expand beyond one school (CCNY) to serve immigrant students nationwide.
  

How We Built It

The project was built with a modern full-stack setup:

• Frontend: Next.js (React) with TailwindCSS for a clean, responsive UI.
  
• Backend: Supabase for authentication and storage.
  
• AI Layer: LangChain to orchestrate prompts, memory, and retrieval.
  
• Data Crawling: Firecrawl to ingest and index scholarship and resource pages.
  
• Deployment: Vercel for frontend hosting and API routes.
  

A typical query flows through the LangChain pipeline, which parses the question, retrieves relevant resources, and generates a structured response for the user.

Challenges We Faced

• Debugging LangChain & Firecrawl: Getting both to work together took significant effort, especially when parsing inconsistent HTML pages.
  
• Resource Constraints: Managing API credits and hosting costs while still experimenting quickly.
  
• Output Refinement: Early versions of the agent produced overly verbose or unstructured responses, so designing prompts and formatting rules became a major focus.
  
• Time Management: Building an end-to-end working prototype in a hackathon timeframe while ensuring stability was tough but rewarding.
  

Built With

• css
• fastapi
• firecrawl
• langchain
• postgresql
• python
• react
• tailwind
• typescript
• vercel