VAIBHAV TANWAR

Engineered a scalable, Dockerized MLOps platform utilizing annotation enabled continual learning to enhance wildlife monitoring capabilities using camera trap images, achieving <60ms average API latency across all modules and scalable data handling of over 1M image records for a seamless Human-in-the-Loop annotation experience.

Improved detection accuracy of multi-class endangered species by 54% by finetuning YOLOv8 on proprietary wildlife dataset from Wildlife Institute of India, and cut inference latency by 97% on average using TensorRT quantization, enabling near real-time processing.

Built a scalable Open Set Re-Identification service using MegaDescriptor embeddings and CLIP model, leveraging PostgreSQL with pgvector, for efficient low-latency vector and semantic searches across millions of images.

Designed and implemented Active Learning pipelines to alleviate high labelling cost in species segregation and Bird Count modules by selecting more informative samples to label based on instance level uncertainties, achieving a 3.5% increase in Mean Average Precision and 26.3% decrease in annotation budget.

Curated and refined domain-specific training datasets for supervised finetuning of LLMs in complex reasoning tasks.

Leveraged Reinforcement Learning from Human Feedback to align model predictions with human preferences, enabling iterative improvement in truthfulness, harmlessness, and instruction-following.

Assisted in creating and optimizing reward models by curating preference datasets derived from human feedback.