Harsh Raj
CSE @ IIT Dharwad
I'm a 2022 graduate at Indian Institute Of Technology Dharwad, pursuing B. Tech in Computer Science and Engineering. Being a tech-enthusiast, I love to explore new technologies and leverage it to solve real-life problems. I am driven by the will to create an impact and encourage diversity and inclusion in communities.
I have demonstrated history of being a good algorithmic problem solver, by securing various good ranks on the contests held on Codechef and other recognised competitions like ACM ICPC. Currently I hold a rating of 2238, which evaluates to 6 stars on Codechef.
I am highly interested in building and contibuting to softwares that can positively impact lives of people. Currently, I am exploring the exiting field of Deep Learning ( Computer Vision, Natural Language Processing and Deep Reinforcement Learning ). Would love to collaborate on projects deploying ML services to Large scale distributed systems.
Work Experiences
ML Intern
- Worked as an ML Intern in the TB Adherence team that aims to accurately classify the Risk for Loss To follow up of TB patients.
- Major work included enhancing the dataset by encoding the categorical features in a more meaningful way.
AI ENGINEER INTERN
- Surveyed research papers corresponding to promising models like EDSR, DCSCN, RDN etc for the task of Image Super Resolution using deep generative architectures.
- Trained EDSR model on company’s private dataset. Added functionality like Learning Rate Scheduler, Gradient clipping etc using pytorch to add support to training on amazon aws EC2.
- Worked on the task of Person Re-Identification using various Deep Learning models.
- Wrote Scripts to generate the corresponding datasets from company’s private video files, implemented pipeline to finetune and benchmark the performance of various existing models on the new dataset.
DEEP LEARNING INTERN
- Along with otherinterns, performed literature survey and experimented with various Encoder Decoder architecture based Convolutional and Recurrent Neural Network models in orderto come up with an efficient video captioning model.
- Read and summarized various research papers like Inflated 3D ConvNet (I3D), Bidirectional attentive fusion, Reconstruction Network etc.
- Gained hands on experience implementing and training various elementary neural network architectures using LSTM, Resnet etc for classification tasks and fine-tuning heavy models like Bidirectional Attentive Fusion using PyTorch on amazon aws EC2.
Key Projects
Dis Sim
- Built a Microservices architecture based image similarity measuring system.
- Used Fastapi for backend server, Celery and Redis for message queues, and Streamlit for frontend server. Packaging and deployment was done using Docker.
- Added authentication functionality using Jwt tokens, and data monitoring service using Apache Kafka.
- Added Subscription based rate limiting of user's requests using Redis.
- Added continious integration of docker images on docker repository (dockerhub) on new release of the repo using github actions.
- Improved observability of microservices by monitoring using Prometheus and Grafana for Fast Api microservices and Flower for Celery.
- Used multi-staged docker builds to reduce the size of docker containers by 50%.
- Performed load testing using Locust achieving a response rate of 350ms for 99.9% of the requests at an average of 46 requests per second.
Reinforcement Learning
- Explored various model free Reinforcement Learning algorithms under supervision of Prof. Prabuchandran KJ
- Implemented various algorithms like epsilon-greedy, UCB, Thompson Sampling, Reinforce for solving the task of Multi Armed Bandits using Numpy. Studied affect of various parameters on the final regret.
- Implemented value based models like Q-Learning, SARSA, MonteCarlo and policy based models like Proximal Policy Optimization on environments like CartPole-v1, MountainCar from scratch using PyTorch.
Multi Image VQA
- Worked on the multimodal task of Multi Image Visual Question Answering under the supervision of Prof. Prabuchandran K.J.
- Explored various pre-existing works for image feature extraction using top-down and bottom-up attention, and the fusion of image and question features. Created a couple of new, balanced and more robust datasets for the given task.
- Implemented attention based models using pytorch, with different feature extractors and trained with different loss functions to achieve competent accuracy score.
CSE HUB
- Single handedly developed a website that aims to provide platform for Algorithmic problem solving, contest hosting, online code editing using built in editor, discussion forum and much more.
- Used Django framework with sqlite database for back-end, and HTML, JS with bootstrap4 for front-end.
- Used multiprocessing to speed up the evaluation of submitted code, wrote various unit tests and automated them using github workflows, packaged using docker and hosted it on cloud platform.
PYTORCH IMPLEMENTATIONS
- Experimented with various CNN based models and their ensembles in orderto perform multi-class image classification.
- Secured 127th global rank, obtaining a weighted F1 score of 0.910 on the test dataset on HackerEarth’s Deep Learning Challenge.
- Implemented the research paper ”Siamese Neural Networks for One Shot Image Recognition” from scratch. Used Omniglot dataset to train the model using PyTorch achieving competent score.
- Implemented the research paper ”Stacked Attention Networks for Image Question Answering” from scratch using PyTorch.
- Implemented the Tab Transformer architecture proposed in the paper "TabTransformer: Tabular Data Modeling Using Contextual Embeddings" from scratch using PyTorch. Wrote clean and well structured code, following best practices and latest technologies.
- Trained the model on "Blastchar dataset" experimenting with various internals like activation function, type of attention used etc along with different parameters. Studied the differences in the results using analytical reports using Weights and Biases.
ENTITY BASED SENTIMENT ANALYSIS
- Implemented an attention based model in orderto perform Entity Based Sentiment Analysis of the tweets. Used a pre-trained BERT as a feature extractor.
- Our team Bagged 2nd rank in Inter IIT Tech Meet 9.0 providing solution to 2 more NLP tasks including automatic Headline Generation and Mobile-Tech Classification of tweets.
AUTOMATED PAYROLLE
- Collaborated with my 2 colleague to develop website aimed to automate payroll calculation process for firms by tracking the attendance of employees using various techniques.
- I handled query and update of database using MySql, various form validation using Php and JavaScript and fronted styling using bootstrap.
Programming Achievements
ACM ICPC
- My team Oracle was one of the 3767 teams to qualify for ACM ICPC Amritapuri regionals securing 1st rank in college.
6 STAR AT CODECHEF
- I became 2238 rated at codechef, securing 96th rank globally (out of around 13 thousands participants) in Codechef Jan Long Challenge Div1.
SMART INDIA HACKATHON
- Qualified for finalround of Smart India Hackathon, 2020, a national level hackathon organised by Government Of India, witnessing participation of more than 2 lakh students.
GOOGLE HASH CODE
- My team Int elligence secured global rank 1207 out of more than 10.7 thousand teams teams competing globally.
SAMSUNG CODATHON
- My team secured 3nd rank at Samsung codathon, an inter college algorithmic programming competition held at KLE Hubballi.
INTER IIT TECH MEET
- My team received bronze medal at Inter IIT Tech Meet, 2020 providing heuristic solution to BOSCH’s event ofroute optimization algorithm.
CODECHEF LONG CHALLENGE
- I was ranked 18th globally out of around 10 thousands participants and 7th in India in Codechef June Long Challenge Div2.