Muhammad Abdullah Baig

Summary

What I did

• Served models via FastAPI behind NestJS microservices on AWS, with LangChain LLM coaching feedback and an RPE-driven workout recommender.
• Built an event-sourced league platform with Socket.IO live scoring.
• Developed a ~130-endpoint recommender with ~94k products ingested.
• Deployed an ML microservice for RPE recommendation on an in-house RTX 3060.
• Fine-tuned YOLOv8-Pose and YOLO11-Pose on internal footage.
• Fed keypoint coordinates into a LightGBM model to score movement trajectory and timing.
• Used YOLO object detection to track players and balls across frames.
• Owned both Liftbuddy and Athleads products end to end, including infrastructure, backend, deploys, and ML services.

Results

• Replaced a per-exercise matrix with a regression model for RPE and weight prediction that landed within ±1 of user-reported RPE 85-90% of the time.
• Developed a computer vision form analysis model that achieved 92% accuracy under optimal conditions and approximately 80% in real gym environments.
• Implemented on-device key frame extraction and compression to minimize latency and bandwidth for video analysis.
• Integrated a CV model with an LLM to convert form scores into actionable coaching feedback for users.
• Achieved an end-to-end p99 latency of approximately 9 seconds at 50 concurrent users for the form analysis and LLM feedback pipeline.