Pace Copilot

Pace Copilot is built upon the Vercel AI SDK, a framework that facilitates seamless integration with multiple LLM providers. This flexibility enables developers to experiment with various models to evaluate performance in generating structured training plans.

Model Comparisons

Several LLMs were tested for training plan generation, including:

• GPT-4: Delivered highly coherent and contextually relevant plans with superior speed and reliability. Ultimately, it was selected for deployment due to its balance of performance and stability.
• Deepseek (deepseek-chat): Produced high-quality plans with accurate and structured outputs, but exhibited occasional instability. It could be an excellent alternative to OpenAI, with prices 100x cheaper per number of tokens.
• Mistral: Although offering a cost-effective solution, Mistral’s response times were significantly longer, and the generated plans were less refined and inconsistent in quality.

To ensure high-quality, deterministic outputs, the temperature parameter of the model was set to 0.3, balancing creativity with reliability.

Authentication and Credit System with Clerk

To manage user authentication efficiently, Pace Copilot integrates Clerk, which allows for a quick and seamless authentication process. Using Clerk, user accounts are secured, and access control is streamlined. Additionally, to regulate the number of training plans generated per user, a credit-based system was implemented. Each user is allocated three credits, with each plan generation consuming one credit. This system helps maintain fair usage and manage resource allocation effectively.

Generating Personalized Training Plans

Pace Copilot create a custom prompt for the LLM based on the athlete’s input, ensuring the generated training plan is tailored to their specific needs. The application collects key information such as weekly mileage, preferred activities, and target events, which are then used to construct a highly specific prompt.

Workflow

1. User Input Collection: The athlete provides key information such as weekly mileage, preferred activities (road running, trail running, strength training, cycling), and target events.
2. Dynamic Prompt Engineering: The application constructs a highly specific prompt incorporating the user’s data to guide the model toward generating a tailored plan.
3. Training Plan Generation: The LLM processes the structured input and produces a comprehensive, step-by-step training schedule.
4. Output Validation: The generated plan is validated using a Zod schema, ensuring all necessary fields are correctly structured and aligned with training principles.

Schema Validation with Zod

A crucial component of the system is the Zod schema, which enforces a structured format for the model’s output. Each field in the schema is meticulously defined with detailed descriptions, helping the model generate precise and usable data. This validation step ensures the output remains consistent and aligns with evidence-based training methodologies.