Hylite Studio

Stats claim. Video proves. Hylite Studio uses footage as hard data to audit scouting reports, exposing the discrepancies between the numbers and the reality.

Comment

Inspiration

Football is torn between two philosophies: Jose Mourinho’s "eye test" and Moneyball’s "data revolution." Hylite Studio bridges this gap. It acts as an AI auditor that cross-references a player's scouting report (stats/PDFs) against raw game footage to verify if the numbers actually match the tape.

How it uses Gemini 3.0 Pro (Integration)

Hylite leverages the advanced cognitive reasoning of Gemini 3.0 Pro to perform this "audit." I utilized three specific next-generation features:

  1. Deep Semantic Video Understanding: unlike previous models that only recognized surface-level action, Gemini 3.0 Pro analyzes intent. It understands off-the-ball movement and tactical positioning in my video context, allowing it to validate complex claims like "vision" or "defensive awareness" that earlier models missed.
  2. Cognitive Audit Persona: Gemini 3.0 Pro's instruction-following allows for a highly nuanced "Skeptic" persona. I program the model to actively debate the input data against the visual evidence, detecting subtle discrepancies (e.g., stat-padding passes vs. progressive play) with human-level discernment.
  3. Structured Reasoning Output: I utilize the model's native ability to output strict, complex JSON schemas even during heavy reasoning tasks. This powers the frontend dashboard, delivering a precise recruitment priority score (0-10) and frame-perfect evidence timestamps that map the AI's cognitive process directly to the video player.

How I built it

I built Hylite as a solo developer using Go (Golang) on the backend for high-concurrency stream handling and Vue 3 (TypeScript) for the interactive "Case File" dashboard. The application uses a decoupled architecture where the frontend handles instant local playback via Blob URLs, while the backend manages the heavy lifting of reasoning and multimodal context processing via Google's generative AI.

What's next for Hylite Studio

I plan to evolve the system into an "Agentic Jury." Instead of a single AI model deciding the outcome, I want to implement a consensus system where multiple specialized agents (Tactical, Physical, Data) debate a player's performance before delivering a final verdict.

Built With

Share this project:

Updates