Yadira Romancelly
AI Data Specialist & Computer Vision Engineer
Results-driven Computer Vision and Data Operations Engineer with 3+ years of experience at Peloton driving high-accuracy ML data pipelines. Proven track record managing millions of image/video annotations across 100,000+ datasets, boosting model accuracy by 30%, and delivering hundreds of thousands of dollars in operational savings through workflow automation. Expert in human pose estimation, structured labeling schemas, and cross-functional technical workflows.
Technical Skills
Data Operations
Image/Video Annotation
Dataset Curation
Cleaning & Normalization
Augmentation
Edge Case Resolution
Taxonomy Implementation
CV Frameworks
PyTorch
OpenCV
TensorFlow
Libraries & Tools
Python
SQL
NumPy
Pandas
Labelbox
CVAT
SuperAnnotate
Evaluation & Formats
COCO Format
Pascal VOC
mAP
IoU
Precision / Recall
Confusion Matrices
Professional Experience
July 2021 – Present | Remote, USA
- Scalable Data Operations: Engineered and executed high-volume data operations pipelines, annotating and validating millions of high-precision images across 100,000+ datasets to train advanced human pose estimation models.
- Model Optimization: Achieved a 25% performance increase and a 30% boost in tracking accuracy by rigorously analyzing model outputs, flagging misclassifications, and tuning precision, recall, and mAP.
- Cost & Workflow Automation: Developed automated dataset labeling and validation scripts that slashed external vendor reliance, saving hundreds of thousands of dollars in operational overhead.
- Resource & Efficiency Leadership: Scaled team throughput and data integrity by designing lean QA workflows, reducing the required QA analyst headcount from 12 down to 3 while maintaining strict data quality standards.
- Data Preprocessing: Built robust Python-based preprocessing pipelines for dataset cleaning, normalization, and semantic augmentation to optimize training quality.
- Cross-Functional Collaboration: Collaborated with ML research teams to define 1,800+ exercise movements, document complex edge cases, and establish core annotation guidelines.
Key Projects
Peloton Strength Library & Feature Engineering (Product Development)
- Core Feature Development: Formulated the underlying dataset architecture and taxonomy used to define, categorize, and develop the entire Peloton exercise and strength library.
- Form Feedback Engine: Curated and structured high-precision keypoint and bounding box datasets using the COCO format to train real-time, biophysical form feedback algorithms.
- Peloton IQ Integration: Spearheaded model prediction auditing and failure pattern mapping to continuously retrain and refine live automated tracking features.
- Data Pipeline Architecture: Designed core data pipelines and technical labeling schemas for real-time tracking features. ➡️ Review visual assets and disclaimers in my Portfolio Examples.
Education & Certifications
Master's Level Work in Computer Science
University of Kansas (Lawrence, Kansas)
Professional Certifications
Deep Learning Specialization (Coursera / DeepLearning.AI)
Computer Vision with PyTorch
GIS & Remote Sensing Certification