Reformer Pilates App

Business Context

Reformer Pilates is one of the most effective low-impact workouts for core strength, posture, and flexibility — but it’s expensive, inaccessible, and rarely personalized. Studio classes cost between $2,000–$8,000/year, offer limited availability, and often move too quickly for real-time form correction.

I wanted to create a way for people — especially women aged 25–45 — to do Reformer Pilates at home while receiving:

Personalized form correction and cueing
Adaptable workouts based on life stage, injuries, or goals
Expert-level instruction using the equipment they already have

This app bridges the gap between high-quality studio guidance and flexible, affordable, at-home practice.

Technical Approach

I developed a Retrieval-Augmented Generation (RAG) app using LangChain and GPT-4.1-mini. The app answers form-related questions using expert-sourced Pilates content from transcripts and manuals. It’s deployed on Hugging Face via Chainlit, with plans to move to Streamlit or Vercel for improved UX.

Stack Highlights:

LLM: GPT-4.1-mini
Embeddings: Snowflake Arctic (fine-tuned)
Vector Store: FAISS
UI: Chainlit → migrating to Streamlit
Evaluation: RAGAS
Monitoring: WandB

Implementation Challenge

Decision Rationale

Trade-offs and reasoning

Chose FAISS over Pinecone for cost-efficient, local prototyping
Fine-tuned Snowflake Arctic Embed for domain-specific Q&A
Added Tavily API as a fallback for unseen queries
Faced consistent Hugging Face deployment failures post-finetuning — resolved through community duplication
Opted for Chainlit to accelerate demo, moving to Streamlit for production polish