The Contender is a unique, strategy-driven, head-to-head calisthenics application that blends human strength and competitive intelligence, all judged with objective AI precision. It transforms simple bodyweight exercises into a thrilling, real-time sporting event.

The app pits two players against each other in a series of no-equipment, bodyweight exercises (Push-ups, Squats, Planks, etc.).

Format: Best-of-three rounds.

Winning: The player who performs the highest number of repetitions or holds the longest time with perfect form wins the round.

The Key Strategy: The person who loses the current round earns the right to choose the exercise for the next round. This forces players to choose a challenge that targets the opponent's weakness while playing to their own strength.

The application's unique features are driven by two distinct AI technologies working in concert:

The player's own device (webcam or phone camera) acts as the unbiased official, ensuring fairness and accuracy.

• Rep and Form Validation: The CV engine (TensorFlow.js/MediaPipe) analyzes the player's movements in real-time. It only counts a repetition if the body hits the exact angles and depths required for perfect form (e.g., chest within one inch of the floor for a push-up).

• Real-Time Feedback: Players receive instant visual or audio cues (e.g., "Deeper!", "Perfect Rep") if their form falters.

Powered by the Python backend (FastAPI + OpenRouter), the Large Language Model manages the strategy and experience.

• Strategy Advisor: After a round loss, the LLM analyzes the player's strengths/weaknesses and the opponent's performance history. It recommends the most advantageous counter-exercise for the loser to pick to maximize their chances of winning the next round.

• Form Referee: Before each round begins, the LLM generates the strict, official form rules for the chosen exercise, which are then enforced by the CV Judge, removing all human ambiguity.

• Hype Generator / Narrator: The LLM provides dramatic, personalized commentary and post-round analysis (e.g., "The Machine sees fit to bestow the title 'Tricep Terror' upon Person B..."), turning data into an engaging narrative.

• Real-Time Synchronization: Scores update instantly (rep-by-rep) across the network using WebSockets, giving players a true feeling of a neck-and-neck race.

• Dynamic Difficulty: The LLM can introduce adaptive handicaps (e.g., forcing a dominant player to perform a harder variation like a Diamond Push-up) to ensure balanced, highly challenging matches.

• Battle History: Detailed statistics are tracked, including personal records for every exercise, win/loss ratios, and streaks, fueling long-term competitive drive.

• Universal Access: The competition requires zero equipment, meaning battles can take place anywhere using a standard webcam or smartphone.

• Frontend: React.js + TypeScript + Vite + Tailwind CSS
• Backend: FastAPI (Python) + Uvicorn + WebSockets
• CV Judge: MediaPipe/TensorFlow.js (runs locally in browser)
• LLM Intelligence: OpenRouter API (Python SDK)
• Database: Cloud-based database (PostgreSQL)
• Authentication: Firebase Authentication
• Hosting: Vercel (Frontend) + Railway (Backend)
• Multiplayer Sync: WebSockets (Rep-by-Rep)

For technical implementation details, see PLAN.md.