Gemini Scout — Time-Traveling Talent Scout

Discover your Team USA destiny with Gemini Scout. A 5-agent pipeline that uses 120 years of data to match your potential to Elite/Adaptive pathways and time-travel your story across Games history.

Comment

The Journey of Gemini Scout

Inspiration: From Data Charts to Living Legacies

As someone who isn't a die-hard sports fan, I didn't want to build a sea of charts and numbers that felt cold. I wanted a fan-centric experience, something that could live on the Team USA website and make every fan feel like they have a place in the Olympic and Paralympic story. My goal was to move away from "data printouts" and toward a narrative that brings a smile to the user's face.

How I Built It: The Architectural Evolution

This project was a "speed run" in AI engineering. I had never built with agents before, and my journey moved through three distinct architectural phases:

  1. The Trio: Started with a simple 3-agent setup.
  2. The Supervisor: Pivoted to an LLM-led Supervisor to handle dynamic routing.
  3. The Sequential Pipeline: Ultimately moved to a deterministic ADK `SequentialAgent` pipeline to ensure 100% reliability and low latency.

The system is now a sophisticated 6-agent orchestration (Scout, Narrator, Compliance, Logger, Eval, benchmark user impersonator) that balances creative storytelling with strict brand guardrails.

Challenges & Technical Rigor

The biggest challenge was Authenticity. Since I’m not a sports expert, I didn't know if my "Scout" was actually making sense. To solve this, I built a custom Bench Testing System and an Eval Agent.

I spent days fine-tuning thinking budgets and temperatures, watching my benchmark scores climb from $5.5$ to $7.4/10$. Seeing the scores move in real-time as I adjusted the agent instructions was a "lightbulb moment" for me as a developer.

The "Judge's Vault": Safety & Transparency

I want the judges to focus on three things:

  • The Multi-Agent Architecture: A robust pipeline that handles everything from Euclidean biometric matching to time hopping narration.
  • The "Mission Control" Logger: I built this specifically for you! It exposes the "brains" of the agents so you can see the reasoning tokens in plain English.
  • The Compliance Guardrails: I "stress-tested" the system by attempting to badmouth the Games and Google; the agents remained professional and friendly every time.

Paralympic Parity & Archetypes

Because I couldn't use specific athlete names (NIL) or Games trademarks, I developed 14 unique Athletic Archetypes based on 120 years of historical data. This allowed me to give every user a meaningful "Standing" and "Adaptive" pathway of equal depth—ensuring that Paralympic representation is a core feature, not a footnote.

  • Paralympic Benchmarking: I created 8 specialized disability personas (including limb difference, visual impairment, and SCI) for my benchmark system. I ran these through the system to verify that the Scout and Narrator provided the same depth, dignity, and accuracy for Para-sports as they did for standing sports.

What I Learned

This hackathon motivated me to learn AI orchestration from the ground up. I’ve gone from knowing nothing about agents to building a benchmarked, multi-agent system that I am incredibly proud of. This is just the beginning for Gemini Scout!

Built With

Share this project:

Updates