As a Tech Lead, I led a dynamic team of 4 members, overseeing and orchestrating various technical projects. My role involves not only managing the team but also actively engaging in the critical stages of project development. I specialize in conducting Proof of Concepts (POCs) to validate technological feasibility and efficiency. Additionally, I contribute significantly to system design, aligning technical solutions with the specific requirements presented by product owners.
Previously, I was a Research Engineer at DiDi Labs where I architected & implemented a Graph-Based Neural Network to forecast the heading direction of pedestrians within a scene to aid our car's decision-making process. Boosted the model's efficiency from 65% to 72%.
Python
80%
MySQL/DynamoDB
70%
LLMs
50%
what i did.
Machine / Deep Learning
Built smart autonomous agents using deep learning tools and AI planning
Robot Perception
Task Planning
Teaching Assistant
Grader & Teaching Assistant for the course CSE 571: Artificial Intelligence
Developed assignments in ROS and Gazebo
Developed auto-graders for homeworks & assignments
Helped undergrads with AI concepts
recent work.
Here's some of my recent work
all
deep learning
ai
Roblocks: An Educational System for AI Planning and Reasoning
Automated Car Parking Detection
3D Object Reconstruction
Visual Servoing
MNIST DCGAN
Pix2Pix(CGAN)
CycleGAN
Image Segmentation
AI Search using ROS & Gazebo
Project Overview
Roblocks: An Educational System for AI Planning and Reasoning
Architected an end-to-end system which introduces AI planning concepts using mobile manipulator robots. It uses a visual programming interface to make these concepts easier to grasp. Users can get the robot to accomplish desired tasks by dynamically populating puzzle shaped blocks encoding the robot's possible actions. This allows users to carry out navigation, planning and manipulation by connecting blocks instead of writing code.
Devised algorithms that can detect cars parked at multiple parking spots, compare if two cars are same or not, predict the color of a car and output each car that was detected and how long it was parked for (approximately) within a given time interval.
Integrated ROS enabled 3D Recurrent Reconstruction Neural Network (3DR2N2) to generate the 3D shape of an object
from 2D images and detected grasping poses on it.
Utilized a deep object detection network (YOLO) to capture an object’s movements in the current camera frame which then served as evidence to a Partially Observable Markov Decision model for visual servoing.
A tensorflow implementation of DCGAN on MNIST dataset. The model has been only trained for 50 epochs. If you want your model to be really good then you should run it for more epochs.
A tensorflow implementation of Pix2Pix(DCGAN) on CMP Facade dataset. The model has been only trained for 150 epochs. If you want your model to be really good then you should run it for more epochs.
A tensorflow implementation of CycleGAN. For the generator network a modified unet model implementation from tensorflow is used. THe model has been only trained for 50 epochs. If you want your model to be really good then you should run it for more epochs.
A tensorflow implementation of Image Segmentation network. The model being used here is a modified U-Net. In-order to learn robust features, and reduce the number of trainable parameters, a pretrained model for the encoder is used. The encoder is a pretrained MobileNetV2 model, whose intermediate outputs are used, and the decoder consists of the upsample blocks already implemented in TensorFlow Examples in the Pix2pix tutorial. The decoder has been only trained for 20 epochs. If you want your model to be really good then you should run it for more epochs.
Artificial Intelligence is the study of building agents that act rationally. Most of the time, these agents perform some kind of search algorithm in the background in order to achieve their tasks. Typically, there are two categories of search, Uninformed and Informed. In this project, we help our turtlebot3 to find paths through his maze world, avoiding obstacles and reaching the goal location and finally hitting the target object (and only the target object) in red.
This system has been designed and developed with the motivation that it could be used to teach about Search problems in the course of AI.