Peak | Devpost

Inspiration

Athletes have access to an incredible amount of advice from coaches, doctors, and the internet. The problem is that they’re expected to piece it all together themselves: tracking metrics in spreadsheets, switching between apps, and constantly adjusting for sleep, recovery, and injury risk without a unified system.

Most trainers treat these variables in isolation, when as athletes, we understand how difficult it is to find a solution for each symptom. Training should constantly evolve, with training, nutrition, and recovery working together to produce the best results.

What it Does

Peak is a multi-agent AI training and nutrition system built around a chat-first experience. Instead of navigating complex menus, athletes can simply ask:

“What should I eat to hit 180g protein?”
“What should I lift today?”
“My shoulder hurts. What now?”
“How’s my progress looking?”

Behind the scenes, four AI agents collaborate:

1. Nutrition AI

Generates meal plans
Suggests best local food options aligned with macro targets

2. Training AI

Tracks progression (1RM, volume, consistency)
Designs and auto-adjusts workouts based on fatigue or injury

3. Health AI

Performs structured injury assessments
Generates rehab and treatment protocols

4. Mountaineer AI

Combines data from all agents to produce tailored results
Summarizes daily briefings of training, nutrition, and recovery

Combined, these four agents function as:

A performance dashboard
A macro tracker
A lifting analytics tool
A rehab assistant
A conversational performance coach

How We Built It

Frontend

React + TailwindCSS for UI
shadcn/ui for clean components
Mapbox + Google Maps API for macro-aware restaurant search

Backend

FastAPI (Python) for API architecture
PostgreSQL for structured storage

AI Architecture

Peak is designed as a multi-agent system rather than a single LLM model.

Each individual agent has a domain specific prompt and access to relevant data under its category. The Mountaineer agent then routes user queries and injects context from nutrition, training, and health databases to provide a unified response.

Retrieval-Augmented Intelligence (RAG)

To avoid generic LLM advice, we integrated a Retrieval-Augmented Generation (RAG) layer. When the user asks for a prompt, our models curate the top 200 most relevant evidence-based sports recovery, nutrition, and strength training methodologies. This is put into context before generating advice for our athletes, ensuring historically accurate advice for their unique situations.

Challenges We Ran Into

One of our biggest challenges was using all four of our specialized AI agents to work together in favor of the user. We wanted to avoid returning generic LLM responses to each nuanced problem the athlete had, so we had to help the agents pass context between each other, and create a Mountaineer agent to store data and make final decisions.

Accomplishments that We're Proud Of

Built a fully functioning multi-agent AI performance system
Integrated Retrieval-Augmented Generation (RAG) to ground responses in 200+ evidence-based performance methods
Integrated Google Maps API to implement restaurant search
Created a chat-first interface that is intuitive for new users

What We Learned

We learned how to work under pressure, adapt on the fly, and how to build products that cater to specific audiences. On the technical side, we also became more experienced with creating API routes with FastAPI and learned to use RAG to ensure situation-specific advice.