ajay-kumar.com
CV
ajay@portfolio: ~ — zsh
ajay@portfolio:~$

Ajay Kumar, computer scientist

Ph.D. candidate in computer science at the University of Missouri, expected Spring 2027. Research intern at Siemens Technology. Open to Applied Scientist and ML Engineer roles in 2027.

LinkedIn GitHub Google Scholar ORCID CV (PDF)

About

A computer scientist four years into a Ph.D. on making science run faster, with a recent stretch learning how LLM agents fail.

Who I am

I make large scientific workloads run faster and cheaper, and I build the checks that tell you when an AI system is wrong.

I am a computer scientist and Ph.D. candidate at the University of Missouri, working with Dr. Praveen Rao in the Scalable Data Science Lab. My thesis is about optimizing scientific workflows, mostly genomics and neuroimaging pipelines, on NSF cyberinfrastructure such as FABRIC and CloudLab. In practice that means multi-GPU scheduling, Slurm clusters, knowledge graphs over billions of statements, and a lot of benchmarking.

Since July 2026 I have been a research intern at Siemens Technology in Princeton, working on validation for agentic LLM pipelines. Before the Ph.D. I spent two years in statistical genomics at the International Rice Research Institute, where I built DeepMap and trained more than a hundred scientists. I have also shipped three iOS apps of my own.

Ph.D. candidate, 4.0 GPA Ph.D. expected Spring 2027 Vice President, ACM UPE Mizzou
Ajay Kumar speaking at a University of Missouri lectern
Presenting at the University of Missouri
Now

Validating agentic LLM pipelines at Siemens. Two invention disclosures, one provisional patent application, one paper under review at CAIN 2027.

Ph.D. expected Spring 2027

By the numbers
2×speedup over greedy scheduling, GPU variant calling
72%less energy on a CPU-GPU cluster, neuroimaging
5peer-reviewed papers, 2 as first author
100+scientists trained in ML and genomics
Recognition
  • FABRIC Travel Fellowship$5,500
  • UPE Executive Council Award$2,500
  • KNIT9 / MERIF Travel Award$2,000
  • IEEE IISWC Travel Award$1,500
  • Claude Builder Club Hackathon1st place
  • Outstanding Teaching AssistantEECS, 2026
  • Outstanding Ph.D. StudentMizzou
Stack
LLMs & evaluation
Fine-tuningRAGAgentic pipelinesEvaluation harnessesUncertainty & abstentionHugging FaceClaude API
ML & data
PyTorchTensorFlow / Kerasscikit-learnGraph MLApache SparkRDF / SPARQLPandasRSQL
Systems
Multi-GPU training & inferenceNVIDIA ParabricksSlurmNextflowDockerKubernetesFABRICCloudLabAWSAzureGCPSwiftUI
Certifications
  • AWS Certified Cloud Practitioner
  • Responsible Conduct of Research · CITI, 2024
  • Genomic Data Science · Johns Hopkins, 2020
  • AI for Medicine · DeepLearning.AI, 2020
  • Deep Learning · IIT Madras (NPTEL), 2020

Experience

Research, industry and the odd side venture, newest first.

Jul 2026 – presentPrinceton, NJ

Research Intern, Agentic AISiemens Technology · Foundational Technologies · with Dr. Christof Budnik

  • Led applied research on validating a production LLM-to-knowledge-graph pipeline for building and HVAC maintenance data, from an open-ended product need to a working validation framework.
  • Inspected every stage boundary of a full production run, identified eight types of silent failure, and added eight new checks. Three of the failure types came from deterministic code, not from the LLM.
  • Outcomes: two invention disclosures, one US provisional patent application as co-inventor, and a first-author paper under review at CAIN 2027.
agentic AILLM evaluationknowledge graphsPython
Aug 2022 – presentColumbia, MO

Graduate Research & Teaching AssistantScalable Data Science Lab, University of Missouri · advisor Dr. Praveen Rao

  • First author, IEEE IISWC 2026: evaluated the DeepPrep neuroimaging pipeline on 6 public datasets and 4 cluster configurations on NSF FABRIC. A CPU-GPU cluster cut energy by about 72% and power by about 74% compared with a high-core CPU cluster.
  • First author, ACM BCB Companion 2025: combined ML runtime prediction with job-shop scheduling for human-genome pipelines on GPU machines, for an average 2× speedup over a greedy approach and 1.6× over a dynamic approach.
  • Helped build VariantKG: 1,508 human genome sequences turned into about 4 billion RDF statements, with graph machine learning on top (Frontiers in Big Data, 2025).
  • Built and deployed the G2PDeep-v2 web platform for multi-omics phenotype prediction, used by about 120 researchers since 2023.
  • Fine-tuned and benchmarked BERT and RoBERTa on about 22 million COVID-19 vaccine tweets; developed the cyberinfrastructure curriculum (FABRIC, CloudLab, ACCESS, Slurm, Globus) for NSF CyberTraining award OAC-2417875.
distributed systemsGPU computinggenomicsneuroimagingSlurmFABRIC
Aug 2020 – Jul 2022Hyderabad, India

Project Associate I, Statistical GeneticsInternational Rice Research Institute · South Asia Hub

  • Built DeepMap, an open-source deep learning package for genomic prediction in crop breeding that outperformed classical models by 13–20%; published to PyPI and used in about 10 later studies.
  • Led the AWS–IRRI partnership that grew compute capacity for genomics researchers by 50%.
  • Trained 100+ scientists across institutions in R, GWAS, genomic selection and Python automation; co-authored Et-GWAS (Life Science Alliance, 2024).
deep learningGWASRAWS
2020 – 2022Remote

AI Event Ambassador & podcast hostDeepLearning.AI (Pie & AI Hyderabad) · The Scientific Monk on YouTube

  • Hosted 10+ episodes with AI researchers and faculty on getting into Ph.D. programs and industry research, reaching about 10,000 listeners and a community of 8,000+ learners.
Jun – Aug 2019New Delhi, India

Software Engineer InternDesign Innovation Center, School of Planning and Architecture

  • Tested and shipped features for Urban STeP, a government participatory-planning site that reached 15,000 visitors; built photo tagging for a biodiversity field app on Android.
Ph.D. · Expected Spring 2027

Computer Science, University of Missouri

GPA 4.0. Thesis: optimizing scientific workflows on NSF-funded cyberinfrastructure. Advisor: Prof. Praveen Rao.

M.Sc. · 2018 – 2020

Computer Science, Banaras Hindu University

8.23/10. Thesis supervised by Dr. Manoj Kumar Singh.

B.Sc. · 2015 – 2018

Computer Science, University of Delhi

Where the 10,000 hours of coding started.

Research notes · Vol. 5, 2026 Publications & talks Distributed computing · Genomics · AI

Making genome-scale science run on any cluster, and knowing when the model is wrong.

5peer-reviewed
2first author
1under review
17citations

Citations from Google Scholar, October 2026, counting only papers I am an author on.

Fig. 1 · GPU variant calling, average speedup
Over a greedy approach
2×
Over a dynamic approach
1.6×
Human-genome variant calling with ML runtime prediction and job-shop scheduling on GPU machines. ACM BCB Companion 2025.
Fig. 2 · Neuroimaging preprocessing, relative energy
High-core CPU cluster
1.0
CPU-GPU cluster
0.28
Energy use running DeepPrep on six public datasets and four cluster configurations on NSF FABRIC: about 72% less energy and 74% less power on average. IEEE IISWC 2026.

Publications

  1. A Benchmark for Cost and Energy-Efficient Execution of Neuroimaging Workflows on Commodity Clusters

    A. Kumar, V. Omelyusik, S. Nair, P. Rao

    IEEE International Symposium on Workload Characterization (IISWC), Boulder, CO, 2026

    First authorConferenceNSF OAC-2417875
    Abstract

    We evaluate DeepPrep, a GPU-accelerated neuroimaging preprocessing pipeline, on six public datasets and four cluster configurations across makespan, power, energy, carbon and cost. The best configuration depends on the structural-to-functional scan size ratio; a CPU-GPU cluster cut energy by about 72% and power by 74% versus a high-core CPU cluster. We give a procedure for picking the cluster for a chosen target metric. Reproducibility artifact: DeepNeuroBench on GitHub.

  2. Paper on validating a production LLM-to-knowledge-graph pipeline

    A. Kumar and Siemens Technology co-authors

    CAIN 2027, industry track (co-located with ICSE) — under review

    First authorUnder reviewIndustry
    Summary

    A layered validation framework for an agentic LLM pipeline over building and HVAC maintenance data. It catches silent failures at stage boundaries instead of letting incorrect outputs through. Title and full author list to follow once the review concludes.

  3. Optimizing the Variant Calling Pipeline Execution on Human Genomes Using GPU-Enabled Machines

    A. Kumar, P. Rao, P. Sanders

    16th ACM International Conference on Bioinformatics, Computational Biology and Health Informatics (BCB Companion), Philadelphia, 2025

    First authorConference
    Abstract

    Combines machine learning runtime prediction with job-shop scheduling for human-genome variant calling on GPU machines using NVIDIA Parabricks, for an average 2× speedup over a greedy approach and 1.6× over a dynamic approach.

  4. A Scalable Tool for Analyzing Genomic Variants of Humans Using Knowledge Graphs and Graph Machine Learning

    S. Prasanna, A. Kumar, D. Rao, E. J. Simoes, P. Rao

    Frontiers in Big Data, vol. 7, 2025 · 6 citations

    JournalOpen access
    Abstract

    VariantKG turns 1,508 human genome sequences (650 GB of variant-call data) into about 4 billion RDF statements on NSF FABRIC and CloudLab, so cohort questions become SPARQL queries, and adds graph machine learning for downstream inference. An earlier version appeared on arXiv (2407.20879).

  5. A Scalable Tool for Democratizing Variant Calling on Human Genomes Using Commodity Clusters

    K. Shehzad, A. Kumar, M. Schutz, C. Webb, P. Nalela, M. J. Das, P. Rao

    33rd ACM International Conference on Information and Knowledge Management (CIKM), demo track, Boise, 2024 · 1 citation

    ConferenceDemo
    Abstract

    Variant calling on human genomes using Apache Spark and Hadoop on commodity clusters across FABRIC, CloudLab and AWS. Source: MU-Data-Science/GAF.

  6. Extreme Trait GWAS (Et-GWAS): Unraveling Rare Variants in the 3,000 Rice Genome

    N. Gnanapragasam, V. V. Prasanth, K. T. Sundaram, A. Kumar, B. Pahi, et al.

    Life Science Alliance, vol. 7(3), 2024 · 9 citations

    Journal
    Abstract

    An extreme-trait GWAS method for finding rare variants across the 3,000 Rice Genomes panel, enabling precision genomic selection for crop improvement.

  7. DeepMap: A Deep Learning-Based Model with a Four-Line Code for Prediction-Based Breeding in Crops

    A. Kumar, K. T. Sundaram, N. Gnanapragasam, U. M. Singh, K. J. Pranesh, et al.

    bioRxiv preprint, not peer-reviewed, 2023 · 1 citation

    First authorPreprintPyPI package
    Abstract

    A multi-model deep learning package for regression and classification on genomic data, runnable in four lines of Python. Outperforms classical models by 13–20%; used in about 10 later studies.

Talks & workshops

Where the work has been presented. Drag the map.

2026
Poster

Benchmarking neuroimaging workflows on commodity clusters

Research and Creative Activities Forum, University of Missouri · October 28, 2026

2026
Invited talk

My Siemens research internship

EECS Graduate Student Association, University of Missouri · October 23, 2026

2026
Paper talk

A benchmark for cost and energy-efficient neuroimaging workflows on commodity clusters

IEEE IISWC 2026 · Boulder, Colorado · September 29, 2026 · IISWC travel award

2026
Organizer & instructor

NSF workshop: Scientific Workflows for Brain and Behavioral Research

University of Missouri EECS, virtual · August 13–14, 2026 · 81 registrants, hands-on FABRIC sessions

2026
Presentation

Scholarship recipient presentation

UPE Annual Convention · March 27, 2026

2025
Paper talk

Optimizing variant calling on human genomes with GPU-enabled machines

16th ACM BCB · Philadelphia, Pennsylvania · October 2025

2024
Demo

Democratizing variant calling on human genomes using commodity clusters

33rd ACM CIKM · Boise, Idaho · October 2024

2023
Tutorial

G2PDeep-v2: multi-omics deep learning for phenotype prediction

ISMB/ECCB 2023 · Lyon, France

2021
Invited workshop

R, GWAS, QTLseqr, genomic selection and Python automation for researchers

IRRI South Asia Hub · 100+ scientists trained

2021
Podcast

Pie & AI Hyderabad: how to prepare for a Ph.D., for FAANG, for data science

DeepLearning.AI · 10+ episodes, ~10,000 listeners

Projects

Research software, platforms in production, and apps I built because I wanted them to exist.

1st place
Ajay beside a Claude banner at the hackathon
Hackathon · Claude Builder Club, Mizzou2026

Equine Lameness Predictor

A clinical prototype built with Dr. Kevin Keegan of the College of Veterinary Medicine that localizes lameness in horses from sensor gait data and clinical records. Won first place; I am now re-evaluating it with leakage-free, case-level splits.

Live demoGitHub
Siemens Technology2026

Validation for agentic LLM pipelines

Validation for an LLM-to-knowledge-graph pipeline over building and HVAC maintenance data. Found eight types of silent failure at stage boundaries and added eight new checks. Two invention disclosures and a provisional patent application.

Proprietary · paper under review at CAIN 2027
Benchmark · IEEE IISWC 20262025–26

DeepNeuroBench

Runs the DeepPrep neuroimaging pipeline on six datasets and four cluster configurations and measures makespan, power, energy, carbon and cost. A CPU-GPU cluster cut energy by about 72% over a high-core CPU cluster.

GitHub artifact
Scheduling · FABRIC2026

MRIPrepCloud

DeepPrep on a four-node Slurm cluster with GPU sharding, critical-path job priorities and an elastic CPU spill cap. Same outputs, 12% shorter wall time, CPU utilization from 71% to 81%.

GitHub
GPU pipeline · ACM BCB 2025, CIKM 20242024–25

GPU-GAF: variant calling at scale

Variant calling on commodity clusters and GPU machines with NVIDIA Parabricks, plus ML runtime prediction and job-shop scheduling for an average 2× speedup over a greedy approach. Tooling from the NSF RAPID project.

GitHubPaper
Knowledge graph · Frontiers in Big Data2023–24

VariantKG

1,508 human genome sequences as about 4 billion RDF statements, so cohort questions become SPARQL queries instead of batch jobs. Graph ML on top for inference.

PaperarXiv
Platform · in production2022–

G2PDeep-v2

Web platform for multi-omics phenotype prediction and biomarker discovery. Python, TensorFlow/Keras, React and SQL; used by about 120 researchers since 2023. I built and deployed the platform and presented its tutorial at ISMB/ECCB 2023.

g2pdeep.org
Package · IRRI2020–22

DeepMap

Open-source deep learning for prediction-based crop breeding: regression and classification on genomic data in four lines of Python. 13–20% over classical models; used in about 10 later studies.

PreprintPyPI
NLP · big data2024

Vaccine sentiment with BERT and RoBERTa

Fine-tuned and benchmarked BERT and RoBERTa for sentiment analysis of about 22 million COVID-19 vaccine tweets.

GitHub
iOS · App Store2026

Operation Fit: Plan & Track

A fitness planner that turns an intake into a nutrition and training plan, then tracks it. SwiftUI, SwiftData, on-device Apple Foundation Models for coaching text, no accounts, no server.

App StoreSupport
iOS · App Store2026

CoMo Events

Every public event in Columbia, MO on one map, fed by a Python aggregator that rebuilds the feed on a schedule with GitHub Actions. Free, open source.

App Storecomoevent.comGitHub
iOS · App Store2026

StackDrop: Tower Stacker

A one-tap block-stacking arcade game in SpriteKit with a seeded daily challenge and Game Center leaderboards. Free, no ads.

App StoreGame page

Earlier work: Urban STeP, a participatory planning toolkit for the School of Planning and Architecture · a biodiversity field app for Android · GestureHandRecognition, hand-gesture recognition · IRRI-Automation · Slurm cluster install scripts · FABRIC slice creator notebooks.

Teaching

Outstanding Teaching Assistant Award, University of Missouri.

Courses · University of Missouri · 2025–26
CS 4080
Parallel Programming for High Performance Computing
multi-process, multi-threaded, distributed and GPU programming
CS 3910
Advanced Cybersecurity
network security, intrusion detection, forensics
CMP_SC 1300
Computing with Data in Python
programming and data science for non-CS majors
CS/ECE 4001
Integrated Cyberinfrastructure, also CMP_SC 7001/8001
neuroscience pipelines on FABRIC and CloudLab
CS 8001
Graduate teaching materials
cloud computing and Python
Workshops & curriculum
  • Developed the cyberinfrastructure curriculum (FABRIC, CloudLab, ACCESS, Slurm, Globus) for NSF CyberTraining award OAC-2417875 — cyneuro.github.io
  • Co-organizer and instructor, NSF workshop "Scientific Workflows for Brain and Behavioral Research", August 13–14, 2026, with 81 registrants
  • Trained 100+ scientists at IRRI in R, GWAS, QTLseqr, genomic selection and Python automation
  • Student mentor for project teams across all four University of Missouri System campuses, 2025–26

Reviewing & service

Peer review, committees, judging and the communities I help run.

Peer review
  • Artifact Evaluation CommitteeACM/IEEE CGO 2027
  • ReviewerAMIA Annual Symposium 2024 (24 reviews) · NeurIPS (ethics review) · ACM ICMR · Frontiers in Artificial Intelligence and other Frontiers journals
  • Sub-reviewerIEEE MDM 2024 · ACM/IEEE JCDL 2024 · ICWE 2024 · IEEE BIBM 2023
  • ACM Certified ReviewerAssociation for Computing Machinery, 2026
Judging & mentoring
  • JudgeMizzou Engineering High School Research Competition, 2026
  • JudgeShow Me Research Week poster session, 2026
  • JudgeTigerX Hackathon, University of Missouri, 2025
  • Technical mentorEntrepreneur Quest student startup competition, Fall 2026
Leadership & community
  • Vice President, Upsilon Pi Epsilon, Mizzou chapterSince Aug 2024, re-elected for 2025–26. Started a research showcase series, brought in speakers from NVIDIA and Netflix, and organized resume reviews with hiring managers.
  • AI Event Ambassador & podcast host, DeepLearning.AI2021–22. Hosted 10+ episodes with AI researchers and faculty, reaching about 10,000 listeners.
  • Community Lead, Community AI HubFree AI education for learners worldwide, with a focus on reaching women.
  • Volunteer and ambassadorIndia Redefined, a non-profit · Student Involvement Ambassador, University of Missouri.
  • MembershipsUpsilon Pi Epsilon (lifetime) · ACM · IEEE · American Society of Human Genetics.

In the news

Coverage from Mizzou Engineering, followed by a running log of updates.

Mizzou EngineeringFeb 16, 2026

Computer science students take on neuroscience

Students in the NSF-funded cyberinfrastructure course presented projects from EEG image reconstruction to MRI depression research. Ajay and Vladimir Omelyusik supported the teams with neuroscience pipelines and cyberinfrastructure.

Read the article →
Mizzou EngineeringJan 16, 2025

Increasing access to cybertraining technology

Satish Nair and Praveen Rao received a nearly $500,000 NSF grant to launch an advanced cyberinfrastructure training program for neuroscience. Ajay is pictured with the team and developed the curriculum.

Read the article →
Updates
Nov 2026

Joined the CGO 2027 Artifact Evaluation Committee.

Oct 2026

Presented a poster at the University of Missouri Research and Creative Activities Forum.

Oct 2026

US provisional patent application filed as co-inventor, from my work at Siemens.

Oct 2026

Invited talk on my Siemens internship for the EECS Graduate Student Association.

Oct 2026

Operation Fit, my first iOS app, went live on the App Store. StackDrop and CoMo Events followed.

Sep 2026

Presented at IEEE IISWC 2026 in Boulder, supported by an IISWC travel award.

Aug 2026

Co-organized and taught the NSF workshop "Scientific Workflows for Brain and Behavioral Research", with 81 registrants.

Jul 2026

Joined Siemens Technology in Princeton as a research intern.

Mar 2026

Presented as a scholarship recipient at the UPE Annual Convention.

2026

First place at the Claude Builder Club Hackathon at Mizzou. Outstanding Teaching Assistant Award, Department of EECS. ACM Certified Reviewer.

Dec 2025

$2,500 UPE Executive Council Award from the Upsilon Pi Epsilon International Honor Society.

Oct 2025

Presented GPU-enabled variant calling at the 16th ACM BCB in Philadelphia.

2025

Paper published in Frontiers in Big Data: VariantKG, knowledge graphs and graph ML over human genomic variants.

Oct 2024

Demo at ACM CIKM 2024 in Boise: democratizing variant calling on commodity clusters.

2024

$5,500 FABRIC Travel Fellowship and $2,000 KNIT9 / MERIF Travel Award.

2024

Et-GWAS published in Life Science Alliance.

What colleagues say

Words from people I have worked with.

His innovative work on the G2PDeep application significantly improved AI system performance. Ajay's speed, accuracy, and clear communication make him invaluable. Any organization would be fortunate to have Ajay on their team.

Trinath AdusumilliFull Stack Developer, Mastercard

Ajay Kumar is an exceptional computational biologist and programmer with expertise in AI, ML, DL, and software development. His proficiency in developing accurate packages and software for biological data analysis is impressive.

Bandana PahiBioinformatician, IRRI South Asia Hub

Ajay Kumar excels in Computational Biology and AI, demonstrating eminent expertise and a unique approach to exploring new methods.

Vamshi VenkatResearch Associate / Data Scientist, Johns Hopkins University

Having known Ajay since college and worked with him on various projects, I can attest to his hard work and commitment to excellence in researching new domains and applying ML/DL concepts.

Atul JoshiData Scientist, UnitedHealth Group

In my collaborations with Ajay, I have found him to be sincere, hard-working and dedicated. I'm sure he will continue to excel wherever he goes.

Varad SrivastavaGenerative AI, Barclays
Community

7,700+ followers on LinkedIn, 10+ podcast episodes with AI researchers, and 100+ scientists trained.

LinkedIn YouTube GitHub

Contact

Hiring for 2027, or want to collaborate on a paper? Write to me.

Email

ajay.ducs@gmail.com

LinkedIn GitHub Scholar ORCID X CV (PDF)
Send a message