FitScout – Project Story

Inspiration

As a parent of high school athletes navigating the college admissions and recruiting process, and I saw how frustrating and challenging it can be. There wasn't a place for student-athlete to get a clear answer to the question:
"Can I get into this school and can I make the team?"

The information is fragmented across admissions websites, athletics portals, NCAA data, and spreadsheets. I wanted something my kids could use and something that could eventually scale to help others with the same challenge.

What it does

FitScout is an LLM powered college research assistant for student-athletes.
It takes in a simple student profile, including academic metrics and athletic performance (race PRs) and a target college. Actions:

• Searches official admissions and athletics sources
• Extracts structured and unstructured data (acceptance rate, tuition, team roster, recruiting standards, etc.)
• Fills in missing gaps with targeted follow-up queries
• Generates a personalized report : comparing the student to the school’s academic profile and athletic benchmarks

The output is a realistic, data-driven fit assessment built specifically for the student-athlete.

How we built it

We used Google’s Agent Development Kit (ADK) to design a multi-agent system, leveraging:

• SequentialAgent to orchestrate the pipeline
• LlmAgent to handle individual responsibilities like search, data extraction, and report writing
• google_search as a tool for real-time web queries
• pydantic models to enforce structure for academic and athletic data

The agent workflow includes:

1. A university information search agent
   
2. A sports data search and extraction agent
   
3. A unified missing information resolver
   
4. A final report-writing agent

The model backbone is Gemini-Pro, configured to run on Vertex AI when deployed in the cloud.

Challenges we ran into

• Extracting consistent athletic data across conferences and event formats (especially in track and field)
• Prompting agents to reliably adhere to schema-based outputs
• Managing N/A fields by ensuring agents could follow up appropriately without hallucination
• Balancing depth of research with runtime performance to keep the system fast
• Dealing with nuances of Vertex AI configuration and environment specific behavior

Accomplishments that we're proud of

• Built a fully working MVP using structured multi-agent orchestration in ADK
• Achieved high schema accuracy and data relevance in the web extractions
• Designed the app with real-world users in mind and it’s something my own kids could actually use today

What we learned

• The value of agent specialization: having small, focused LLM agents perform better than general-purpose ones
• Schema-driven prompts improve reliability and enable downstream reasoning
• Real-world tasks like college matching require both factual extraction and comparative reasoning; LLMs are surprisingly effective at both using good prompts
• ADK is a easy to use framework for orchestrating autonomous agent pipelines using the ADK search tool.

What's next for FitScout

• Add support for multi-college comparisons across academics, athletics, and cost
• Build a UI so students and parents can interact with the system more easily
• Expand beyond track and field into sports like swimming, soccer, volleyball, and basketball
• Enable profile persistence and progress tracking over time
• Integrate outreach tools to connect with college coaches directly

Built With

• google-adk
• python
• vertex