Srijith Rajamohan, Ph.D.

AI Lead

Washington, DC

Education

2009-2014 Ph.D. in Computational Engineering The University of Tennessee

2007-2009 MS in Electrical Engineering The Pennsylvania State University

Leadership

• Lead product and research data science teams
• Develop the vision and long-term strategy for applied data science
• Lead and develop a cross-product research portfolio
• Mentored data scientists from graduate to principal levels

Core Competencies

Experience (selected)

2022-Present Staff AI Research Scientist, Sage AI

2022-2022 Senior Data Scientist, NerdWallet

2020-2022 Senior Developer Advocate (Data Science & ML), Databricks

2014-2020 Computational Scientist, Virginia Tech

2009-2014 Graduate Research Assistant, SimCenter: Center of Excellence in Computational Engineering, UT

Staff AI Research Scientist

I currently lead the data science and research portfolio within product teams at Sage AI. My responsibilities include finding novel solutions to difficult problems by interfacing cross-collaboratively across teams with Product Managers, Data Scientists, ML Engineers and business stakeholders to solve novel business problems.

Links: = Webpage of project

Projects

Thought leadership

• - Developed and led the research portfolio for developing an accounting-specific LLM
• - 7 patent submissions
• - Authored the strategy document ‘Applied data science within product teams’
• - Built a prior induced plan-and-act LLM-based framework for decision making ‘Constrained Hierarchical Planner using LLMs`
• - The Road to AGI

Jan 2024 - Present

Tech lead for domain-specific LLM training (Collaboration with AWS)

• - Lead a team of PMs, engineers and data scientists to train an accounting-specific LLM from OSS LLMs
• - Roadmap planning, experimental design, stakeholder management
• - Pretraining and finetuning of OSS LLMs such that they ingest new accounting knowledge and can be used in a variety of product use-cases

July 2023 - Present

Retrieval Augmented Generation (RAG) - LLM for Q&A (Patent filing in progress)

• - Lead the data science effort to build a QA engine for a customer-facing product
• - Architected a robust QA system (from SQL DB) using a redesigned vector store
• - Proposed a novel POS-enhanced retrieval system from vector store for context extraction

June 2023 - Sep 2023

Tech lead for cashflow forecasting

• - Led a team of three to perform cashflow forecasting from invoice data
• - Identify patterns in invoice history for short term forecasts of Accounts Payable data
• - Built a forecasting endpoint and a Dash dashboard

Mar 2023 - Present

Generative AI - LLM-based model for response generation (Patent pending)

• - Lead the data science effort to deploy OpenAI ChatGPT for intent detection and response generation
• - Proposed a method to improve the accuracy of model (intents) by 20%
• - Identify and quantify certainty of hallucinated responses (text-davinci)
• - Built a model monitoring notebook

Mar 2023 - Jun 2023

Intent detection from text using classical ML

• - Built and deployed an XGBoost classifier to predict intent from text
• - Compare the results of XGBoost classifier with zero shot ChatGPT classifier
• - Improved model accuracy using active learning and incremental learning for small data

Feb 2023 - Mar 2023

Applied Reinforcement Learning for Financial Applications

• - Overview of RL (including RLHF) algorithms presented to the Sage AI team
• - Identifying opportunities for applying RL within financial accounting
• - Evaluation of RLLib for online learning and D3RLPY for offline learning

September 2022 - Feb 2022

Identifying Recurring Trusted Transactions from vendor buyer interactions (Patent pending)

• - Identify transactions that are recurring and can be classified as trusted from vendor buyer transactions.
• - This allows the users of Sage products to help automate payments
• - Interact with stakeholders, identify and translate the problem, create a research proposal
• - Architect an ML framework using unsupervised density estimation to solve the above problem
• - Setup Kedro pipelines for repeatable DS experimentation

September 2022 - October 2022 Graph Neural Networks for Financial Applications

• - Presentation on using GNNs for various applications in financial accounting such as fraud detection, trusted transactions etc.

Sr. Data Scientist

In the data science team at Nerdwallet, I design experiments and perform modeling to solve data science problems for business units such as lead optimization. I also interface with the data engineering and ML infrastructure teams to optimize ML workflows.

Links: = Webpage of project

Projects

April 2022 - June 2022 Contextual bandits for product placement/ordering

• - Designed a contextual bandit-based framework to address the cold-start problem in product placement and recommendation.
• - This takes as input a context vector corresponding to a class of users and learns the appropriate ordering for that class over time
• - Evaluated contextual versions of algorithms such as epsilon-greedy, Thompson sampling, LinUCB etc.

May 2022 - July 2022 User understanding for prospective and propensity models

• - Maintenance of a ML/DL pipeline in for propensity and prospect models from user data
• - User data obtained from clickstream and transunion data populated from Snowflake
• - Unsupervised autoencoder-based model for generating user vectors
• - Vectors are fed to supervised models (XGBoost and Neural Network) for both propsect and propensity modeling
• - Pipeline was scheduled and automated using Airflow DAGs

April 2022 - July 2022 ML monitoring for production pipelines

• - Architected a proposal for monitoring ML models that are in production (using Evidently.ai)
• - Endpoints consist of both AWS Cloudwatch and Datadog

June 2022 - Jul 2022 ML job submission and workflow management

• - Design a solution for the ML experimentation and production platform
• - Command-line tools designed to keep track of and query submitted jobs, metadata and results
• - Read logs from AWS Cloudwatch using the awslogs Python package
• - Kedro for data science experiments using the PyData stack
• - Data lineage and provenance of generated data
• - Systematic tracking and visualization of experiments

Sr. Developer Advocate (Data science/Machine learning)

My role as a Developer Advocate in data science allows me to serve as a thought leader in machine learning and data science, and educate the community about the state-of-the-art. This role also allows me to engage in internal advocacy, and work cross-functionally across various units such as product management, product marketing, solutions and engineering. Some of my responsibilities include:

• - Thought leadership articles and presentations on enterprise and open-source ML/Data science
• - Provide guidance and feedback to product management and product marketing
• - Act as a subject matter expert to the solutions and engineering team

My technical areas of expertise here are Deep Learning for Natural Language Understanding (NLU), Bayesian inference and large-scale processing using PySpark.

Links: = Webpage of project

Projects

2021 Lead the DevRel efforts on Machine Learning at Databricks

• - Engage with all stakeholders, i.e. everyone from the executive team to the practitioner community to drive adoption of ML on Databricks
• - Guide the ML/DS Product Management team with regards to product features
• - Work with the Product Marketing team to help reach the appropriate community of practitioners
• - Provide OSS solutions for ML on the Databricks platform

2021 Apache Spark

• - Lead the efforts on the growth and messaging architecture of Spark and Koalas
• - Engage with C-suite executives to define the growth strategy for Spark
• - Offer strategic guidance on improvements to the Spark website and documentation, and grow community adoption

2021 Lead the advocacy efforts on OSS MLflow

• - Interface with the MLflow product team at Databricks and drive adoption of OSS MLflow

2020-2021 Authored & open-sourced a set of courses ‘Introduction to Computational Statistics for Data Scientists’

• - A practical guide to getting started with scalable Bayesian Inference using PyMC3
• - Introduction to Bayesian Statistics
• - Bayesian Inference with MCMC
• - PyMC3 for Bayesian Modeling and Inference

2020-2022 Articles on Data Science and Machine Learning

• - GPU-accelerated Sentiment Analysis Using Pytorch and Huggingface on Databricks
• - Are GPUs really expensive? A benchmark study for inference in NLP
• - How wrong is your model?
• - An Experimentation Pipeline for Extracting Topics From Text Data Using PySpark
• - MLflow for Bayesian Experiment Tracking
• - Bayesian Modeling of the Temporal Dynamics of COVID-19 using PyMC3
• - Using Bayesian Hierarchical Models to Infer the Disease Parameters of COVID-19
• - Beyond LDA: Dive into BigARTM for Topic Modeling
• - The Modern Chief Data Officer: Transitioning From Defense to Offense
• - Reproduce Anything: Machine Learning meets Data Lakehouse

2020-2022 Presentations/Talks
Machine Learning at Scale (Nov/Aug 2021)

• - Best practices for the full lifecycle of ML projects along with issues such as reproducibility, explainability, trustworth AI and governance
• - Keynote at the IEEE IDSTA conference
• - Invited talk at the ADBIS workshop

Deep Learning at Scale at Databricks (Oct 2021)

• - Presentation on how to scale Deep learning workloads at Databricks at the Big Data Symposium in South Korea

MLflow for the ML Lifecycle (Nov 2021)

• - Presentation on how to use OSS MLflow for model management, reproducibility with MLflow projects and Model registry, and model inference/serving

Bayesian Modeling of the Temporal Dynamics of COVID-19 using PyMC3 at the Data+AI Summit (Nov 2020)

• - Bayesian inference to estimate the disease parameters of COVID-19 from real case data with PyMC3

Presented ‘Maintable HPC and Data Science with Python/C++’ at the National Center for Atmospheric Research (NCAR) (Mar 2021)

• - Scale data science codes written in Python using a JIT approach with Numba
• - Offload compute-intensive portions of the Python code using Eigen and Xtensor in C++

Computational Scientist

This role as a Computational Scientist involved providing Scientific Computing expertise, enabling High-Performance Computing and Visualization solutions and performing research in Machine Learning.

Links: = Webpage of project

Research Projects

2019-2020 Interactive Network Analysis of Social Graphs

• - Network analysis of social media network of political figures
• - Created a distributed deep learning sentiment analysis pipeline with Huggingface Roberta embeddings
• - Generated networks using Graphtools and interactive visualizations of these networks with Sigma.js
• - Network data obtained from Twitter, ETL with PySpark
• - Metabase as a BI dashboard for SMEs

2018-2020 Determining political affiliation from short texts using stance detection(NLP)

• - Created a deep neural network architecture for stance extraction from a weakly-supervised classifier
  using contextual embeddings (Elmo/BERT) to determine political affiliation
• - Fixed embedding generation frameworks such as Doc2Vec and Fasttext were also evaluated
• - Created an interactive web-based visualization tool for stance visualization
  • - Self-attention for model interpretability
  • - Tweets were stored and preprocessed in MongoDB
  • - Python RQ for data acquisition
  • - PySpark and Spacy used for corpus cleaning

2019-2020 Deep Learning lead for the project ‘Eye Gaze tracking for Surgical Training’

• - Oversaw a team of 5 attempting to learn gaze patterns of resident surgeons using Deep Learning
• - Collaboration between the ISE department at Virginia Tech and the Carilion School of Medicine

2019-2020 Co-PI on Jefferson National Lab funded project ‘Next-generation Visual Analysis Workspace for Multidimensional Nuclear Femtography Data’

• - Visual analytics for nuclear femtography data.
• - Big Data analytics and visualization: Investigating various novel ways to enable understanding
  of nuclear physics phenomena with 3 dimensional visualizations.

2015-2020 General Dynamics Collaboration with the Discovery Analytics Center(VT)

• - Computational statistics and dimensionality reduction using unsupervised techniques such as PCA, MDS
• - Weighted Multi-Dimensional Scaling (WMDS) for semantic interaction
• - Formulated a highly-accurate Inverse MDS algorithm
• - Formulated and implemented the optimization scheme for the solution of Inverse MDS in Python using the NLopt optimization package
• - Parallelized Inverse MDS Python code using Numba
• - Accelerated Inverse MDS C++ code for low latency with Eigen

2019-2020 Generative Methods for Stance Detection and Visualization

• - Investigate Variational Autoencoders for extracting stance
• - Use a semi-supervised learning approach with partially labeled data

2019-2019 Reinforcement learning for Eye Tracking in Laparoscopic Surgery

• - Reinforcement learning (ADNet) for learning eye gaze patterns in surgeons during laparoscopic surgery
• - Used to train and evaluate residents

2019 Cost analysis of On-premise Cloud vs. Public Cloud for Virginia Tech

• - Performed a cost analysis of Virginia Tech’s on-premise cloud and compared it to those offered
  by third-party external cloud providers for scientific computing
• - Analyze the tradeoffs of an on-premise cloud for the research community that led to the
  creation of an interactive report used to estimate and compare costs

2018-2019 ICAT SEAD Grant 2018

• - Co-PI on ICAT SEAD grant to analyze and visualize the health and nutrition policies across countries
• - Interdisciplinary collaboration with the Health and Nutrition, Business and SOVA departments
• - Architected an open-source interactive visual query framework

2018-2019 Scheduling and Visualization Application for Idaho National Lab

• - Built a Django and D3 based scheduling and visualization tool
• - Tool helps the Idaho National Lab manage outage tasks for the nuclear power plants
• - Managed and mentored two students in this development project

2016-2017 HNFE Project for Visualization of National Food and Beverage Endorsements

• - Produced interactive web-based visualizations for exploratory analysis of Food and Beverage endorsements for the HNFE department
• - Work was presented to policy makers to understand the impact of celebrity endorsements
• - Produced a web-based analytics framework for visual querying of endorsement data
• - Framework extended using Dask for analyzing Big Data

2016-2017 Interactive 3D Visualization for Nuclear Reactor Pool

• - Multi-user user interactive visualization using open-source technology (X3D) in web browsers
• - This involved collaborating with the Nuclear Engineering department to interface with their
  RAPID toolset to generate a visualization and querying tool.

Infrastructure projects

Cloud Computing

• - Played a key role in architecting the cloud computing infrastructure for research computing using Openstack at Virginia Tech
• - Involved in designing and setting up cloud computing best practices and architectures using OpenStack, LXC and Docker containers

Scientific Visualization

• - ParaView and Visit for large-scale scientific visualization through remote rendering on supercomputers

Collaborative Computing Platforms

• - Architected collaborative-computing platforms for data science and ML using JupyterHub.

Miscellaneous

• • Research member of several NSF funded projects
• • Virginia Tech campus champion for XSEDE, representative for ACI-REF
• • Poster, and paper reviewer and session chair at XSEDE 2015-2018
• • Mentored and supervised several Graduate Research Assistants

Graduate School

Select research work conducted during my Masters and Ph.D. programs.

Graduate Research Assistant

2009-2014 Experience developing and maintaining a parallel 3D time domain electromagnetic solver using the finite element method for open/closed boundary problems such as radar cross sections, waveguides, frequency-dependent materials etc.

• - Distributed code using MPI that can execute over hundreds of nodes
• - GPU offloading using CUDA for acceleration

Graduate Student

2007-2009 At the Microsystems Design Lab at the Pennsylvania State University, I implemented a neural network for skin-tone detection on the IBM Cell and which resulted in a 23x speedup.

Teaching

2018-2019 SuperComputing 2018-209 short talks : Talks@VT series

• - Introduction to Generative Modeling
• - Cloud Cost Comparison: On-premise vs. External Vendor
• - Deep Learning on GPUs with PyTorch for Text Analysis
• - AutoML: An Overview of Automated Machine Learning

2016 - 2018 Lectures and workshops at Virginia Tech

• - Taught undergraduate class ‘CS1064: Introduction to Python’ in the Spring 2016 semester*
• - Taught ‘Text Summarization with Word Embeddings using PyTorch’ for CS4984/5984 in Fall 2018
• - Taught ’Introduction to OpenACC’ lecture for undergraduate class ‘CMDA 3634: Comp Sci Foundations for CMDA’*
• - Introduction to Scientific Python: 150 min handson workshop
• - Introduction to Data Visualization with Plot.ly: 150 min handson workshop

2015-2019 Select seminar classes taught at Virginia Tech

• - Taught workshop titled ’Introduction to ARC Cloud using OpenStack for Machine Learning’*
• - Introduction to Data Visualization
• - Introduction to CUDA
• - TensorFlow for Machine Learning
• - Dask for Out-of-Core Computing: Big Data solutions on your laptop
• - Unsupervised Machine Learning using Sckit-learn and TensorFlow
• - Supervised Machine Learning using Sckit-learn and TensorFlow
• - Deep Learning using TensorFlow and Keras

2015-2018 XSEDE conference workshops

• - A Data Scientist’s Python Toolbox
• - Workshop titled ’Introduction to Machine Learning’ taught at PEARC18
• - Introduction to Scientific Computing using Python
• - Python Pandas for Data Analytics

Publications

*- Chaitanya S. Kulkarni, Tianzi Wang, Nathan Lau, Jacob Hartman-Kenzler, Sarah E. Parker, Srijith Rajamohan, Laura E. Barnes, Shawn D. Safford Applying Deep Learning to Provide Eye Gaze Guidance for the Peg Transfer Task, Jan 1 2021, 16th Academic Surgical Congress

- Srijith Rajamohan, Robert Settlage Informing the On/Off-prem Cloud Discussion in Higher Education, PEARC20, ACM, Portland

- Robert Settlage, Srijith Rajamohan Enabling AI/DL Workloads on HPC Infrastructure through Containers and Open OnDemand. HPCKP20, High-Performance Computing Knowledge Meeting, Barcelona, July 2020

- Rincón-Gallardo Patiño, Sofía, Srijith Rajamohan, Kathleen Meaney, Eloise Coupey, Elena Serrano, Valisa E. Hedrick, Fabio da Silva Gomes, Nicholas Polys, and Vivica Kraak. Development of a Responsible Policy Index to Improve Statutory and Self-Regulatory Policies that Protect Children’s Diet and Health in the America’s Region. International Journal of Environmental Research and Public Health 17, no. 2 (2020): 495.

- Robert Settlage, Srijith Rajamohan, Kevin Lahmers2, Alan Chalker3, Eric Franz3, Steve Gallo4, David Hudak3. Portals for Interactive Steering of HPC Workflows. Nov 2019, Third Workshop on Interactive High-Performance Computing, SC19

- Srijith Rajamohan, Alana Romanella, Amit Ramesh. A Weakly-Supervised Attention-based Visualization Tool for Assessing Political Affiliation. Aug 2019, arXiv:1908.02282 [cs.CL], https:// arxiv.org/abs/1908.02282

- Zhou, M., Rajamohan, S., Hedrick, V., Rincón-Gallardo Patiño, S., Abidi, F., Polys, N., & Kraak, V. (2019). Mapping the Celebrity Endorsement of Branded Food and Beverage Products and Marketing Campaigns in the United States, 1990–2017 ,International journal of environmental research and public health 16.19 (2019): 3743

- Valerio Mascolino, Alireza Haghighat, Nicholas Polys, Nathan J. Roskoff, and Srijith Rajamohan. 2019. A Collaborative Virtual Reality System (VRS) with X3D Visualization for RAPID, The 24th International Conference on 3D Web Technology (Web3D ’19), ACM, New York, NY, USA, 1-8.

- Srijith Rajamohan and Faiz Abidi, Web-based Visualization and Querying of Food and Beverage Endorsements by Celebrities, PEARC19, ACM, Chicago

- Rajamohan, S., Romanella, A., Ramesh, A., A Human-in-the-Loop Deep Learning Based Document Tagging for Stance Detection, CHCI 2019: Algorithms that make you think, Blacksburg.

- Rajamohan,S. and Anderson, W.K. A Modified Streamline Upwind/Petrov-Galerkin Stabilization Matrix for Time-Domain FEM, ACES 2018, Denver

- Rajamohan,S. and Anderson, W.K. Using an Approximate Streamline Upwind/Petrov-Galerkin Stabilization Matrix for the Solution of Maxwell’s Equations in Dispersive Materials, ACES 2018, Denver.

- Abidi, F., Polys, N., Rajamohan, S., Arsenault, L., Mohammed, A. (2018, April). Remote high performance visualization of big data for immersive science. In Proceedings of the High Performance Computing Symposium (p. 5). Society for Computer Simulation International.

- Zhou M, Kraak VI, Rajamohan S, Abidi F, Polys N. Mapping the Celebrity Marketing of Branded Food and Beverage Products in the United States: Policy Implications and Research Needs. 15th World Congress on Public Health. April 3-7, 2017. Melbourne, Victoria, Australia

- Nicholas Polys, Ayat Mohammed, Jagathshree Iyer, Peter Radics, Faiz Abidi, Lance Arsenault, and Srijith Rajamohan. Immersive Analytics: Crossing the Gulfs with High-Performance Visualization. IEEE VR 2016 Workshop on Immersive Analytics

- Rajamohan,S and Anderson, W.K , HPC for Legacy EM Code, a Mixed Language Approach using CUDA. Applied Computational Electromagnetic Society 2012, Volume: GPU for CEM.

- Porting Algorithms to the IBM Cell Processor - an FFT case study. Penn State Research Symposium 2009.