Arpit Sahni Email & Phone Number

Santa Monica, California, United States

Machine Learning Engineer@Generative AI, Efficient ML. * Research,high performance training and inference(large-scale parallelized distributed as well for low power, low memory low compute (low SWAP) enviornments like mobile phones and glasses), data-generation/ curation and deployment of multimodal generative models, (image-image, text to image (RGB.

Pittsburgh, Pennsylvania, United States

Machine Learning

Pittsburgh, Pennsylvania, United States

• Worked on creating a Computer vision pipeline with object segmentation, pose estimation, and point cloud completion in a construction environment for enabling shape-aware path planning.
• Implemented real-time 3-D point cloud completion for line-of-sight occluded regions of vehicles by training a Graph Attention Neural Network based GAN on partial point clouds obtained from combining RGB, depth and.
• Performed model compression by developing a custom student network architecture and distillation training scheme to enable 30% faster inference time and 45% lesser model parameters than teacher network allowing.
• Developed fast and reliable Task Agnostic technique for Uncertainty estimation using Bayesian Networks for achieving Temporal consistency in Video sequence predictions for safer real-time deployment of Deep Neural.
• Developed a Temporal attention Stochastic Transformer to Improve Predicate, Object and Subject Recall for semantic scene graph (Augmented Reality) predictions by ~5% on 3DSSG dataset.

Philadelphia, Pennsylvania, United States

• Designed Chatbot to support technician for on-site product installation. used Retriever-Reader Chatbot Architecture; Dense Passage Retriever (Finetuned) and (LLM) ChatGPT - Reader. Reduced Product installation &.
• Enhanced chatbot assistant with Few-shot Image generation feature (Generative AI) by finetuning Stable Diffusion v1.5 with LoRA adapters and training Controlnets to provide useful visual troubleshooting guidance to.
• Built computer-vision pipeline for product installation verification with Instance segmentation using Mask-RCNN. Achieved AP@IOU=0.5 95%. Reduced data Labelling time by 80% by building SAM (Segment Anything Model).
• Improved robustness with connectivity analysis using Depth-Transformer to Reduce False positives by 95%.
• Proposed a new Knowledge Distillation technique leveraging MiniGPT-4 (LLM, VLM, VQA) and COT prompting to Outperform state-of-art in downstream tasks of object detection & Image classification by up to 7%. (Multi Modal.

• Collaborated with cross-functional teams to build Linear Regression Model to estimate Ticket closing time with R-squared of 0.88.
• Reduced maintenance Costs by developing Time-series model (LSTM) to predict component life & optimize schedules.
• Pioneered an Automatic Visual Inspection system by fine-tuning EfficientNet (CNN) model on a dataset curated from historicaltest-reports; Data Augmentation & Grid-Search to get strong recall of 0.94 and led to.
• Mathematical & Machine Learning models for failure analysis and development for automotive components.
• Developed intuitive GUIs and interactive dashboards for comprehensive analysis of Real-World Usage Profile data obtained from Clutch development trials. Additionally, facilitated post-processing of simulation results.
• Enhanced efficiency by crafting Bash scripts that automated simulation job setup on clusters and simplified result retrieval, ensuring a unified and streamlined simulation workflow.

Bengaluru Area, India

Mathematical and predictive(ML) models

Master Of Science, 4.0/4.0

TA for Machine Learning ** Coursework**: 1. Introduction to Machine Learning (Flagship Course) 2. Computer Vision (PhD Level) 3..

Bachelor Of Technology

Activities and Societies: Student clubs: BasketballRanked 4/250 in Final Year