Utkarsh Jain Email & Phone Number

San Francisco, Ca, Us

San Jose, California, United States

Core Recommendation team.

Palo Alto, California, United States

Worked on addressing hallucinations in Large Language Models with post-editing models. The correction model I developed improved the factuality rate of multiple industry-leading LLMs by 10-15%. Additionally, I optimized the workflow to significantly reduce memory usage and cut inference latency by 87%.Read more on the link below.

Advisors: Prof. Gary Cottrell, Prof. Virginia De Sa, Prof. Taylor Berg-kirkpatrickWorked on building a more brain-like CNN at Gary's Unbelievable Research Unit (GURU). In short, this is achieved by using neural data to constrain the representations learned by the CNN which improves its performance on downstream tasks and against adversarial attacks.This project lies in the intersection of Cognitive Science and Machine Learning.

Advisor: Prof. Melissa GymrekI worked on developing a haplotype-based fine-mapping tool that uses conditional regression models to pinpoint the sets of variants in a cohort of related individuals that are the most predictive of certain phenotypes.

Advisors: Prof. Kelly Frazer, Prof. Graham McVickerJoined Frazer Lab in the School of Medicine. My research focuses on using machine learning to predict the effects of genetic variations on enhancers and promoters. A journey into an unfamiliar territory of Machine Learning in Epigenomics.This is a collaboration project with the McVicker Lab at the Salk Institute of Biological Studies.

TA for CSE 100: Advanced Data Structures (Prof. Debashis Sahoo)

San Diego, California, United States

My research focuses on using machine learning to predict the effects of genetic variations on enhancers and promoters. Working under Prof. Kelly Frazer and Prof. Graham McVicker.This is a collaboration project between the McVicker Lab (Salk Institute of Biological Studies) and the Frazer Lab (School of Medicine, UC San Diego).

Guwahati, Assam, India

I worked on using Natural Langauge Processing and Graph Neural Networks in predicting the home locations of Twitter users using their tweets, profile metadata, and social network graph. This research has significant implications for various online services, such as targeted advertising, opinion mining, and event detection. Advised by Prof. Sanasam Ranbir Singh.1. Built Bi-LSTM and BERT baseline models in PyTorch and Python to predict the home location of Twitter users and achieved an accuracy of 36% and a Mean Absolute Error (MAE) of 703 by using users’ tweets.2. Increased accuracy to 58% and reduced MAE to 516 by incorporating tweet metadata and user-mention network and using field-level Attention layers and Transformer-encoder for feature fusion.3. Developed a loss function to capture hierarchical relationships among geolocations which improved accuracy by 7%.

This was my first formal experience in Machine Learning. The goal of this project was to explore the usage of Reinforcement Learning in the image classification task and compare its efficacy with existing Supervised approaches. Advised by Prof. Aditya Nigam.1. Designed a Reinforcement Learning (RL) based Image Classification pipeline in TensorFlow and attained an accuracy of 92% on the MNIST dataset by using Dueling Deep Q-Network. Increased accuracy by 3% with Proximal Policy Optimization.2. Reduced training time by 40% by implementing Classification with Costly Features training procedure.3. Conducted evaluations on various benchmark datasets and observed competitive performance wrt. supervised methods.In a nutshell, we concluded that RL can be in fact used for image classification but demands significantly longer training time and is more complex to implement as compared to a simple, yet very effective, CNN model.

Teaching assistant for the course CS-307: System Practicum

Newcastle Upon Tyne, England, United Kingdom

This was my first research experience and I worked on using statistical and machine learning approaches for solar nowcasting. Advised by Prof. Mousa Marzband.1. Explored various auto-regressive and exponential smoothing models to forecast solar power generation.2. Worked on Statistical Time-Series Models, and an ensemble of LSTM Encoder-Decoder to predict univariate and multivariate time series.3. Used Approximate Bayesian Computation coupled with MCMC to build non-linear univariate and bivariate time series models. Achieved a Mean Average Percentage Error of 17%.

Saket, Delhi, India

1. Improved Shadow Analysis software runtime by 15% by optimizing the data processing pipeline and migrating CPU-intensive jobs over to GPU with CUDA and Python.2. Implemented a stochastic disaggregation procedure for generating synthetic sets of hourly solar irradiation values, suitable for use in solar simulation design work. Achieved a mean percentage error of 40%.3. Proposed and developed a computationally inexpensive model for synthetic data generation which performed better than previously published works and brought down the average error to 5%. Resulted in a 35% increase in customer satisfaction and annual savings of $100,000 in outsourcing costs.4. Integrated and tested the model with the main production software for client use.

Espoo, Southern Finland, Finland

Teaching Assistant for the course CS-C3150: Software Engineering

Espoo, Southern Finland, Finland

Teaching Assistant for the course CS-C3150: Software Engineering

Mandi

Teaching assistant for the course CS-207: Applied Data Practicum

Chandigarh Area, India

My first internship experience as a Software Developer. The main aim was to automate the data collection process from several metrology tools in the production lab. This data holds importance in the later stages of fabrication where it is used for process control and tool health monitoring.1. Used SECS/GEM protocol for equipment-to-host data communications to automate the data submission procedure into the database to accelerate wafer production.2. Created databases, data entry systems, web forms, and other applications for diverse uses by engineers.

Master Of Science - Ms, Computer Science, 3.97/4.0

Highlighted Coursework: Advanced Data-Driven Text Mining, Machine Learning: Neural Networks and Pattern Recognition, Probabilistic.

Bachelor Of Technology - Btech, Computer Science, 8.56/10.0

Highlighted Coursework: Linear Algebra

Computer Science, 4.24/5.0

Semester Exchange Program. Highlighted Coursework: Machine Learning, Deep Learning, Data Science, Stochastic Processes, Optimization.