Pranjal Vithlani

About Pranjal Vithlani

Experienced Computer Vision enthusiast and researcher with both academic and industry…

Experience

1. Applied Scientist

   Amazon

   Jun 2022 — Present

   Prime Video - Scene UnderstandingFoundational model finetuningImproved the precision by 70% for movie scene’s content descriptor prediction. Fine Tuned a 7b parameter Vision-Language model on prime video dataset for domain alignment and instruction fine tuned the task for maturity rating prediction and content descriptor prediction. Language Modeling for Content ModerationImplemented and productionized the first of multi-level profane content detection in textual modality. Accurately detecting the foul content that is nsfw, not-safe-for-kids or all the maturity rating level, this reduced 75% of operators time to review a movie for maturity rating.Applied the model on 5 special rated territories, which have specific rules, policies and tolerance to foul content. Used LLMs along with an ensemble of models for achieving the performance expected as most of the LLMs have guardrails to such type of content.Fully Automated foul language detection, with zero human intervention, and achieved accuracy of 90%. Long-Form Video UnderstandingAchieved 7% improvement from SOTA on internal prime video audio description dataset and 2% improvement on ActivityNet dataset, to work with Longer time length videos, specifically we were able to scale the models to 3 minutes long videos. Published work to AMLC 2024Weakly Supervised Instance SegmentationIntroduced the weakly supervised method based on the data needed to quickly come up with a short term solution. Driving end-to-end projects from data collection, creating labeling tasks, defining metrics for success, training and evaluating models, and Delivery. Reaching true positive IOU till 83% in a quick solution, which was closer to long term goal of 85%Detecting corner points for certain quadrilateral shaped objects, for usage in placement of other type of objects