My Work

A collection of projects spanning computer vision, deep learning, and AI engineering.

High Performance Computer Vision

Real-time tracking solution for customer behavioral and site security analysis with high focus on reidentification.

Developed real-time multi-camera tracking solution across synchronized cameras for secure site monitoring
Enhanced inter-camera joins using pose triangulation and appearance embeddings
Domain Generalizable object re-identification pipeline on challenging open-set benchmark (40% → 74% mAP)
In-house built tools for Multi-Camera Labeling, Visualization and Analysis

High Performance Scalable

A very low-latency face & license-plate anonymization with direct integration to Milestone VMS.

Implement low-latency face and license plate detection software to anonymize RTSP streams
Integrated into Milestone VMS via Milestone AI Bridge
Improved load balancing to ensure high GPU utilization using shared decoders and batching.

High Performance Scalable

Real-time 3D reconstruction & contour fitting using depth cameras and Open3D.

Multi-process real-time pipe & filter architecture for 3D Reconstruction, pointcloud cleaning and contour fitting.

High Performance Scalable

Kubernetes based model serving, benchmarking & safety evaluation for ML products.

Implement serving and testing of PyTorch models over Kubernetes pods.
Designed benchmark to evaluate internal defect predictors on real data as well as synthetically introduced defects
Designed 4 new safety performance indicators to measure safety of autonomous vehicles

Scalable

PostgreSQL/PostGIS platform for curated dataset storage, querying & sampling.

Development of PostgreSQL-based inventory database with geo-coordinate support for world map datasets
Django based backend to manage user requests with celery based workers for high compute tasks
Client library for sampling datasets from the database

High Performance

Single framework for classification, detection & segmentation with 90%+ GPU utilization.

Development of single pipeline supporting Image Classification, Object Detection and Semantic segmentation tasks
Re-implementation of CPU intensive metrics using multi-processing and cython to remove any wait for GPU
Optimization of training and inference phase with resulting GPU utilization of over 90%

High Performance Scalable

Cost-efficient media compression, metadata analysis & beautification using OpenCV/FFmpeg.

Optimizations to efficiently use servers, leading to 70% cost reduction
Enable dynamic compression of media using SSIM metric
Quantifying social posts ownership by measuring plagiarism using image metadata analysis
Implemented Beauty filter based on variational approach by Farbman, Zeev, et al

High Performance Full-Stack Dev

High-performance DynamoDB workflow improving completion rate 5× with 99.5% payload reduction

Development of data model to reduce user interaction for form completion and data payload for mobile usage
Improvement in form completion rate by 5x.
Reduction of data payload by 99.5%.

Scalable Machine Learning

Spark based clustering system reducing labeling time from weeks to under an hour.

Built a platform for QA team to analyze and correct incorrectly labelled products
Utilized unsupervised learning to cluster Amazon catalog data with Human-in-the-Loop analysis
Reduced catalog quality check time from several weeks (previously infeasible) to under an hour

High Performance

GPU-accelerated non-rigid registration achieving massive real-time speedups.

Development of Atmosphere Turbulence Removal Module based on Non-Rigid Registration method
Optimizations added using Pipe-And-Filter architecture and cuda implementation, achieving 240x speedup

High Performance

Real-time stitching & object tracking using CCD/IR cameras with PTZ control.