Repsense

Repsense is an AI-powered fitness companion that turns workout history into adaptive programming. It combines a FastAPI backend, an agentic LLM pipeline, and a Next.js frontend to generate routines, capture feedback, and iterate on training plans.

Key Features

• Routine generation based on uploaded workout history
• Feedback capture with routine resolution and episodic memory
• Routine history with contextual metadata and quick “Add to chat”
• Magic authentication for secured routine access
• Turso-backed storage with a lightweight SQL migration system

Architecture

• frontend/ — Next.js app (Magic auth, chat UI, routines, history)
• backend/ — FastAPI API (profiles, chat, routines)
• agentic/ — LLM pipeline and intent/feedback logic

Requirements

• Python 3.9+
• Node.js 18+
• Turso database (libsql)

Environment Variables

Backend (/Users/prakharojha/Desktop/me/personal/repsense/.env):

• OPENAI_API_KEY — required for LLM calls
• TURSO_DATABASE_URL — Turso connection URL
• TURSO_AUTH_TOKEN — Turso auth token (if required)
• MAGIC_SECRET_KEY — Magic admin secret for token validation
• OPIK_ENABLED — optional (1 to enable tracing)

Frontend (/Users/prakharojha/Desktop/me/personal/repsense/frontend/.env.local):

• NEXT_PUBLIC_API_URL — backend URL (default http://localhost:8000)
• NEXT_PUBLIC_MAGIC_PUBLISHABLE_KEY — Magic publishable key

Setup

Backend

cd /Users/prakharojha/Desktop/me/personal/repsense
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Run database reset and migrations (Turso):

python3 -m backend.storage.reset_db

Start the API:

uvicorn backend.app:app --reload

Frontend

cd /Users/prakharojha/Desktop/me/personal/repsense/frontend
npm install
npm run dev

Frontend routes:

• http://localhost:3000/ — Login
• http://localhost:3000/home — Landing page
• http://localhost:3000/routines — Routine history

How it works

1. Users upload their workout history (CSV).
2. The system generates a profile and uses it to create routines on request.
3. When users give feedback (positive, negative, or injury), the system resolves which routine they mean and stores the outcome as episodic memory.
4. Future prompts are enriched with recent chat context and past routine outcomes to generate better recommendations.
5. Users can open new chats and still refer to previously generated workouts to improve and provide feedback to the agent.

Evaluation and observability

We integrated Opik tracing to monitor LLM calls and enriched query construction, validating when memory and feedback are used in generation and evaluating output quality.

Routine Feedback Resolution

Feedback is resolved via a single LLM call that compares the user’s feedback text against the five most recent routine candidates. If ambiguous, the system asks a clarification question.

Migrations

Migrations are stored in:

• /Users/prakharojha/Desktop/me/personal/repsense/backend/storage/migrations

The reset script drops all tables and re-applies migrations:

python3 -m backend.storage.reset_db

Notes

• Routine endpoints are protected with Magic auth. If MAGIC_SECRET_KEY is missing, routine fetches will fail.
• Routine history titles are disambiguated in the UI when duplicates exist.