Joseph Arun

BrainCorp

• Map Merging : generating one route from multiple routes
• Visual SLAM : stand alone slam system deployed on a small fleet based on monocular vision
• Path Planning : Auto path planning and area filling

The Cottrell Lab, Graduate Student Researcher

• Speeding up natural products discovery using Deep Learning (SMART)
• Aiding doctors in pulmonary heart disease detection from MRI, CT and Echo-cardiogram data

Texas Instruments (TI), Analog Design Engineer

• In charge of designing and validating high precision SAR ADCs
• Developed ultra pure sine wave generators with self-calibration for testing high precision analog parts at less than 1/100 th the current price (link paper)
• Modeled device variations due to fabrication irregularities and estimated product yield and spec stability

Sattva MedTech, Co-Founder

• Co-founded a company in medical devices space (Sattva).
• Selected among top 10 tech innovations in 2015 by Stanford Business School and FICCI at IIGP 2015.
• Developed a low-cost portable solution for fetal monitoring in rural India

Small Molecule Accurate Recognition Technology (SMART) to Enhance Natural Products Research

• Worked on enhancing SMART project
• Enhanced clustering accuracy by augmenting the contrastive loss with Coord-Conv based auto-encoder loss
• Worked on mixture decomposition using capsule nets for better clustering results

Dynamic Switching Controllers for varying environmental dynamics

• Created a generalized framework for augmenting model free algorithms for dealing with time-varying model environmental dynamics
• Multiple model ensembles were pre-trained on various environmental dynamics each specilizing on detecting a specific variational class
• Designed a Dynamic Switching controller that detected the ideal controller based on the current environmental dynamics
• Awarded runner up for best poster award among 46 posters presented (link to poster) (link to paper)

Refining Optical Flow estimation making use of semantic information

• Created a new scalable parallel neural network architecture for estimation of optical flow
• Refined the optical flow estimation by jointly predicting semantic segmentation and using the loss as a regularizer
• Boosted training speed and overall accuracy by training individual parallel networks with custom loss sensitive only for a specific semantic class (link to paper)

Multi-agent Reinforcement techniques in pong and tic-tac-toe

• Trained two agents to compete against each other in pong and nxn tic-tac-toe
• Created game engines for pong and tic-tac-toe that can be used for multi-agent reinforcement learning tasks
• Reward functions were modified to make agents compete or cooperate in the game of pong
• Various techniques like policy gradient, DQN and actor-critic networks were compared in terms of time and efficiency for both pong and tic-tac-toe (link to report)

Fake News Classification using neural networks and ensemble methods

• Created a News Classifier using MLP assisted with real-time Google News search with overall accuracy of 86% and BER of 0.13
• Created a new dataset for fake new classification called Beautiful Liar using web scrapping from sites like politifcat.com
• Compared various model ensembles for the task (link to report)

Recommender System for Sparse Datasets

• Created a Recommender System that suggests stores to users based on Google Local Business Database
• Designed a system that predicts visits using a hybrid collaborative filter optimized to enhance performance in sparse datasets which obtained an accuracy score of 88.41%
• Using Latent Factor model designed a rating prediction system that had an RMSE of 0.74549

University of California, San Diego

GPA: 3.7/4

Skills

• Visual SLAM
• Deep Learning
• Reinforcement Learning
• Computer Vision
• Machine Learning
• Recommender Systems

Languages

• Python
• C/C++
• R
• GO

Packages and Frameworks

• PyTorch
• TensorFlow
• Keras
• Gym
• Pandas
• OpenCV
• SciKit-learn
• git
• Numpy