• Led development of next-gen search experience leveraging agentic LLMs, Retrieval-Augmented Generation (RAG), and multi-step reasoning orchestration

• Tech lead for computer vision and multimodal research

• Designing and optimizing generative models to enhance product image quality through advanced image synthesis and restoration techniques.

• Developing multimodal large-scale models to accurately predict product attributes and identify mismatches and policy non-compliance in listings.

• Working with petabyte-size video dataset and deploying scalable computer vision solutions in the production environment (PyTorch, AWS Sagemaker, S3, EFS, relational DB, and container registry)

• Created ETL and staging pipeline for measuring the performance of ML models at scale (Kubeflow, Prefect, Kubernetes, EKS, AWS Batch, Lambda, and Terraform)

• Developed a real-time visual search engine dashboard and API service for exploring millions of ultrasound images using KNN and vector database (Amazon OpenSearch, Beanstalk, Plotly)

• Setting up a workflow for model experiment tracking and model unit-testing using Data Version Control (DVC), Weights&Biases, and Kolena platforms

• Developed production-level onboard C++ module to detect stuck circumstances

• Scripted unit tests, checked the module performance on simulation, bench, and field vehicle, and finally deployed the module for the release

• Led the engineering team with agile practices to build a robot for vineyard winter pruning

and pilot it at the customer’s site

The projects that have been accomplished under my supervision:

• Utilized Google Cloud Platform (GCP) for model training and prediction

• Simulated a virtual vineyard in Blender and generated synthetic data for pre-training and

data augmentation

• Applied supervised transfer learning and retrained Mask-RCNN, U-NET, and MMRotate for

segmentation of highly overlapping branches (PyTorch)

• Filtered and manipulated pointclouds in Open3D, registered pointclouds from multiple

viewpoints using ICP and RANSAC optimization

• Evaluated the accuracy of stereo depth estimation from OpenCV block matching, HITNet,

and CREstereo on the synthetic dataset

• Simulated robot’s manipulation in ROS and benchmarked +10 trajectory planners from

MoveIt package

• Integrated, calibrated, and tested ZED2i, RealSense D435i, Analog Device ToF, Occipital, and

Leopard WDR Camera for outdoor sensing and data record

• Developed React-based teleoperation dashboard and utilized roslibjs and WebRTC to

stream real-time video from the robot and transfer messages between the robot and teleop

dashboard through WebSocket and P2P communication

• Successfully demonstrated a prototype of an autonomous security robot in a real environment (parking lot) for potential customers and investors

• Developed a real-time ROS node for license plate recognition

• Tested Facebook’s action recognition deep network (SlowFast) for suspicious behavior recognition on the collected data from the field

• Implemented speech-to-text feature to detect suspicious keywords inside the conversations

• Applied HDL Graph SLAM and created 3D map and localized robot using 3D lidar, wheel odometry

• Tuned global and local ROS planners for autonomous navigation of the robot

• Experimented and evaluated Autoware’s vector map for autonomous navigation inside the parking lots

• Distributed computation between edge devices and running real-time object detection (Yolo), license plate recognition network, and navigation stack in parallel

• Set up Webviz for remote monitoring of the robot