◦

Built scalable data ingestion pipelines on Google Cloud (GCS + BigQuery) for clinical analytics, replacing prototype pandas/SQL workflows

◦

Developed ML-driven pose estimation models for yoga posture correction, generating decision-support insights for physiotherapists

◦

Built a 4D reconstruction pipeline from multi-camera traffic footage using DUSt3R and Metric3D, handling temporal synchronization and corrupted frame detection across diverse camera setups

◦

Designed a YOLOv10-based 3D object detection and pose estimation pipeline, achieving 98% accuracy on cars and 90% on pedestrians across 5TB of annotated data

◦

Developed a depth-conditioned image augmentation pipeline using Stable Diffusion, ControlNet-depth, and Depth Anything to generate geometrically consistent synthetic driving data

◦

Designed and built custom LED array hardware for near-field photometric stereo, targeting 3D surface reconstruction for medical imaging

◦

Trained and evaluated deep learning models on endoscopic prototype captures, reducing mean angular error to 17° and depth error to 0.25mm

◦

Implemented physics-based light source calibration from a 24 LED to a 6-LED photometric stereo setup

◦

Built production ETL pipelines for document processing (OCR extraction, clustering, automated retraining), improving throughput by 25%

◦

Developed a deep learning pipeline in Keras for extracting structured components from hand-drawn forms, achieving 15% accuracy improvement