Skip to content

Gautham Manuru Prabhu

Research Associate, MiCoSys Lab, San José State University · Software Engineer 2, AI Acceleration, Cisco

I work on temporal graph learning for graphs whose interactions carry text, such as email threads and event records. At MiCoSys Lab I am testing whether language models improve future-link prediction on these graphs, and which part of the model any improvement comes from.

Interests·Temporal & Dynamic Graphs · Graph Representation Learning · Language Models on Graphs · Parameter-Efficient Learning · Reproducibility in ML

Gautham Manuru Prabhu
Bengaluru, India
peer-reviewed publications
6
citations (Google Scholar)
60+
first-author papers
2
research groups · 3 institutions
4

Updates

News

  • 2026

    Manuscript in preparation with MiCoSys Lab on when language-model semantics help temporal link prediction.

  • Aug 2025

    Promoted to Software Engineer 2 on the AI Acceleration team at Cisco, 12 months after the last promotion.

  • 2025

    Joined the MiCoSys Lab at San José State University as a Research Associate, working with Dr. Saptarishi Sengupta on temporal graph neural networks.

  • Aug 2024

    Converted from intern to Software Engineer 1 at Cisco, six months after joining.

  • 2024

    Graduated from Manipal Institute of Technology with a B.Tech in Computer Science & Engineering and a minor in Big Data Analytics (8.91/10 CGPA, top 15%).

  • 2024

    Placed third of 100+ entries in the Cisco Intern Case Study Competition with an NLP pipeline for standardizing 50,000+ supplier names.

  • Apr 2024

    Our paper on vaccine misinformation spreading over user-association graphs appeared in Procedia Computer Science (ICMLDE 2023).

  • Jan 2024

    Started as a Software Engineering Intern in Supply Chain Operations at Cisco, Bengaluru.

  • Dec 2023

    QuCardio, on quantum machine learning for cardiovascular disease detection, published in IEEE Access (Q1). It has since passed 60 citations.

  • 2023

    Presented SatelTensor at the Tensor Computation & Machine Learning Workshop, IISc Bengaluru.

  • 2023

    Selected for the ACM Winter School on Optimization for ML & OR at IIT Goa, and the Summer School on Dynamic Resource Allocation at the Center for Networked Intelligence, IISc Bengaluru.

  • Aug 2023

    Finished a deep learning research internship at the Medical Informatics Lab, IIT Kharagpur, and presented the OCT anomaly-detection work at the Digital Health Symposium.

  • 2023

    VIKAS and EyeEncrypt, both presented at ATIS 2022, published in Springer CCIS vol. 1804.

  • 2022

    QuCardio reached the Grand Finale of the QETCI Global Quantum Hackathon, top 16 of 1,600+ teams from 25+ countries.

  • 2022

    Led a team of 6 to the Smart India Hackathon Grand Finale (NDRF track) with VIKAS, a multimodal disaster-response system.

Research

Research statement

Most of my research asks what a model gains from structure it would otherwise ignore, and what it costs to use it. The structure has changed from project to project: the geometry of ECG signals, the spread of posts across a social graph, and now the timing of interactions in a graph where every edge carries text. So has the budget. In QuCardio it was counted in qubits. In production systems at Cisco it is counted in milliseconds. At MiCoSys it is counted in trainable parameters.

§1Early work

My first projects were about structure that standard pipelines discard. In QuCardio we tested whether quantum feature maps could separate ECG images that classical kernels could not. QSVC, Pegasos QSVC and a quanvolutional neural network reached 97% accuracy, 10 to 14 points above classical baselines we trained on the same data. SatelTensor approached the same idea with classical tools, using Tucker and CP decompositions to compress satellite image stacks while keeping their spatial and temporal factors separate. Both projects treated the geometry of the data as something to model directly.

§2Learning under constraints

The medical imaging and NLP projects added constraints of a different kind: few labels, noisy inputs, unbalanced classes and, in clinical work, errors that do not cost the same in both directions. At the IIT Kharagpur Medical Informatics Lab I built an ensemble feature-fusion method for anomaly detection in optical coherence tomography scans. At Manipal I built retinal vessel segmentation pipelines over 3,000 fundus images. In the vaccine misinformation project we stopped classifying posts in isolation and modeled how they spread across the graph of user interactions, running BERT and XLNet over 10,000 posts alongside that propagation graph. It was the first project where I modeled the graph itself.

§3Language and time

At MiCoSys, with Dr. Saptarishi Sengupta, I work on temporal text-attributed graphs: graphs where each interaction is timestamped and carries text, such as an email or a news event. Temporal graph neural networks model who interacts with whom and when, and recent work reports large gains from adding large language models to them. Those methods change the text representation and the fusion architecture at the same time, so it is hard to tell which one is responsible. I started by building one of these models myself, a frozen language model conditioned on temporal graph embeddings through LoRA adapters. It underperformed a well-trained text-aware temporal GNN, and finding out why became the project. It is now a reproduction and ablation study, and the manuscript is in preparation.

§4Production systems

At Cisco I build LLM-based agents for Supply Chain Operations and own their retrieval quality, evaluation and guardrails. Four agents now run in production and close 35% of incoming support cases with no human in the loop. Running them has made the cost of scale concrete: latency and memory budgets, input distributions that drift, and the distance between a benchmark score and a deployed system.

§5Future directions

Next, I want to work on how language and structure should be combined in models of evolving graphs. That means knowing when the text adds information the interaction history does not already carry, fusing the two without the fusion step doing all the work, and evaluating these models so that a reported gain survives a change of random seed. Cost matters as well, because the methods that fuse text most deeply are also the most expensive to train and run.

Publications

Publications

Six peer-reviewed papers in reverse chronological order. My name is in bold; ★ marks selected work.

  1. [6]2024★

    Addressing Vaccine Misinformation on Social Media by Leveraging Transformers and User Association Dynamics

    C. Rao, G. M. Prabhu, A. R. Kumar, S. Gupta, N. P. Shetty

    Procedia Computer Science, vol. 235, pp. 1803–1813 (ICMLDE 2023) · Conference

    Adds a model of how posts spread over the user-interaction graph to BERT and XLNet text classifiers, and improves on content-only detection.

    Read paper ↗doi:10.1016/j.procs.2024.04.171
  2. [5]2023★

    QuCardio: Application of Quantum Machine Learning for Detection of Cardiovascular Diseases

    S. Prabhu, S. Gupta, G. M. Prabhu, A. V. Dhanuka, K. V. Bhat

    IEEE Access, vol. 11, pp. 136122–136135 · Journal (Q1) · 60+ citations

    Quantum kernel methods (QSVC, Pegasos QSVC) and a quanvolutional neural network for ECG image classification. 97% accuracy, 10 to 14 points above matched classical baselines. Funded by MeitY and AWS.

    Read paper ↗doi:10.1109/ACCESS.2023.3338145
  3. [4]2023

    VIKAS: A Multimodal Framework to Aid in Effective Disaster Management

    G. M. Prabhu, T. Gupta, M. V. Srujan, A. R. Soumya, A. Palorkar, A. Chowdhury

    Springer CCIS (ATIS 2022), vol. 1804 · Conference

    Combines text and image inputs to triage incoming information during disaster response. First author.

    Read paper ↗doi:10.1007/978-981-99-2264-2_22
  4. [3]2023

    EyeEncrypt: A Cyber-Secured Framework for Retinal Image Segmentation

    G. Hegde, S. Gupta, G. M. Prabhu, S. V. Bhandary

    Springer CCIS (ATIS 2022), vol. 1804 · Conference

    Retinal vessel segmentation paired with Diffie–Hellman key exchange and AES-256, so clinical images can leave the hospital network encrypted.

    Read paper ↗doi:10.1007/978-981-99-2264-2_9
  5. [2]2023

    A Systematic Review of Deep Learning Approaches for Vessel Segmentation in Retinal Fundus Images

    G. Hegde, S. Prabhu, S. Gupta, G. M. Prabhu, et al.

    IOP J. Physics: Conference Series, vol. 2571, p. 012021 · Peer-reviewed

    Reviews the deep learning architectures, preprocessing steps and evaluation protocols used for retinal vessel segmentation.

    Read paper ↗doi:10.1088/1742-6596/2571/1/012021
  6. [1]2023

    SatelTensor: Satellite Data Exploration via Tensor Decomposition

    G. M. Prabhu, S. Gupta

    TCML Workshop, IISc Bengaluru, Tensor Computation & ML Workshop · Workshop

    Tucker and CP decompositions for low-rank representations of satellite image stacks that keep spatial and temporal factors separate. First author.

Full list and citation metrics on Google Scholar.

Talks

Talks & presentations

Talks at workshops and symposia, and finalist presentations at national and international hackathons.

  • 2023

    SatelTensor: Satellite Data Exploration via Tensor Decomposition

    Tensor Computation & Machine Learning (TCML) Workshop, IISc Bengaluru · Contributed talk

  • 2023

    Ensemble Feature Fusion for Anomaly Detection in Optical Coherence Tomography

    Digital Health Symposium, IIT Kharagpur · Research presentation

  • 2022

    QuCardio: A Quantum Ecosystem for Cardiovascular Disease Detection

    QETCI Global Quantum Hackathon, Grand Finale · Finalist presentation

  • 2022

    VIKAS: A Multimodal Framework for Disaster Management

    Smart India Hackathon (NDRF track), Grand Finale · Finalist presentation

Research Projects

Selected projects

Motivation, approach and outcome for each project.

Language Models and Temporal Link Prediction

2025 – Present

Manuscript in preparation · MiCoSys Lab

Motivation. Recent methods report large gains from adding language models to temporal graph networks on text-attributed graphs. They change the text representation and the fusion architecture together, so the source of the gain is unclear.

Approach. A reproduction and ablation study on a public dynamic text-attributed graph benchmark. The temporal graph model stays fixed while the way text enters it changes.

Contribution. Manuscript in preparation. Code and evaluation scripts will be released with it.

PyTorch · PyTorch Geometric · DyGLib · LoRA · Language Models

QuCardio: Quantum ML for Cardiovascular Diagnosis

2022 – 2023

Published · IEEE Access

Motivation. Classical kernels plateau on some ECG classification tasks. Quantum feature maps embed inputs in much larger spaces and might separate classes that classical kernels cannot.

Approach. Compared QSVC, Pegasos QSVC and a quanvolutional neural network against classical baselines we trained on the same ECG image data.

Contribution. 97% accuracy, 10 to 14 points above the classical baselines. Published in IEEE Access and funded by MeitY and AWS. Grand Finalist, top 16 of 1,600+ teams.

Qiskit · Python · Quantum ML · scikit-learn

Paper ↗Code ↗

Misinformation Propagation on User-Association Graphs

2022 – 2023

Published · Procedia CS

Motivation. Text alone misses a useful signal: how a post moves through the network of users who share it.

Approach. BERT and XLNet encoders combined with a graph model of propagation over user interactions, on a corpus of 10,000 posts.

Contribution. F1 above 0.90 on the benchmark, ahead of content-only baselines. My first project with an explicit graph model.

BERT · XLNet · PyTorch · Graph Modeling

Paper ↗

SatelTensor: Tensor Decomposition for Satellite Data

2023

Presented · TCML, IISc

Motivation. Satellite image stacks are large, and generic dimensionality reduction throws away the spatial and temporal structure that later tasks depend on.

Approach. Tucker and CP decompositions to get low-rank representations that keep that structure.

Contribution. Presented at the Tensor Computation & Machine Learning Workshop, IISc Bengaluru.

Python · TensorLy · NumPy

Experience

Research & professional experience

Research positions first, then industry work at Cisco.

Research

Research Associate · MiCoSys Lab (Machine Intelligence & Complex Systems), San José State University

2025 – Present

Advisor: Dr. Saptarishi Sengupta · Remote

  • Testing whether language-model semantics improve future-link prediction on temporal text-attributed graphs, and whether the gain comes from the text or from how it is fused with graph structure.
  • Built an adapter-based language model conditioned on temporal graph embeddings. It underperformed a well-trained text-aware baseline, so the work became a reproduction and ablation study of published methods. Manuscript in preparation.

Deep Learning Research Intern · Medical Informatics Lab, IIT Kharagpur

Jun 2023 – Aug 2023

Advisor: Dr. Subhamoy Mandal · Ms. Pragya Gupta · Remote

  • Proposed an ensemble feature-fusion method for detecting anomalies in retinal OCT scans, combining color-space features with tuned preprocessing and augmentation.
  • Contributed to a dataset of 3,000 annotated retinal scans and presented the work at the IIT Kharagpur Digital Health Symposium.

Undergraduate Researcher · Cybersecurity & Quantum Computing Research Group, MIT Manipal

Feb 2022 – Nov 2023

Advisor: Dr. Vivekananda Bhat · Manipal, India

  • Compared QSVC, Pegasos QSVC and a quanvolutional neural network with classical baselines for detecting cardiovascular disease from ECG images. 97% accuracy, 10 to 14 points above the classical models.
  • Funded by MeitY (Govt. of India) and AWS. Grand Finalist, QETCI Global Quantum Hackathon 2022 (top 16 of 1,600+ teams). Published in IEEE Access.

Undergraduate Research Assistant · Biometrics & Software Engineering Group, MIT Manipal

Jul 2022 – Oct 2023

Advisor: Dr. Srikanth Prabhu · Mr. Govardhan Hegde · Manipal, India

  • Built retinal vessel segmentation pipelines over 3,000 fundus images using CLAHE, Gaussian smoothing and edge detection.
  • The images left Kasturba Medical College's network, so the pipeline encrypts them with Diffie–Hellman key exchange and AES-256. Published in Springer CCIS.

Undergraduate Researcher · NLP / Social Computing, MIT Manipal

Sep 2022 – Jul 2023

Advisor: Dr. Nisha P. Shetty · Manipal, India

  • Modeled how vaccine misinformation spreads, combining BERT and XLNet encoders with a propagation graph over 10,000 posts.
  • F1 above 0.90 on the benchmark. Published in Procedia Computer Science (ICMLDE 2023).

Professional

LLM-based agents and ML services in production at Cisco, Supply Chain Operations.

Software Engineer 2, AI Acceleration · Cisco Systems · Supply Chain Operations

Aug 2025 – Present
  • Four agents in production (service-request resolution, knowledge-base indexing, case follow-up, proactive alerting) close 35% of incoming support cases with no human in the loop, cut mean time to resolution by 40% and return roughly 10,000 engineer-hours a quarter.
  • Own the research-to-production path for these agents: retrieval quality, evaluation and guardrails.
  • Built failure-analysis and predictive-maintenance agents for the Quality Transformation Program. They root-caused 70% of previously undiagnosed failures across global manufacturing sites.

Software Engineer 1 · Software Engineering Intern · Cisco Systems · Supply Chain Operations

Jan 2024 – Aug 2025
  • Promoted twice in 18 months: intern to SE1 in 6 months, SE1 to SE2 in 12.
  • Rewrote legacy failure-analysis workflows as event-driven microservices (FastAPI, Cassandra, Redis on Kubernetes) and added ML anomaly detection over manufacturing telemetry.
  • Placed third of 100+ entries in Cisco's intern case study competition with an NLP pipeline that standardized supplier names across 50,000 records.

Education

B.Tech, Computer Science & Engineering · Manipal Institute of Technology

2020 – 2024

Minor in Big Data Analytics · 8.91 / 10 CGPA · Top 15% of cohort

Recognition

Awards & leadership

Competitions, scholarships, research schools and student leadership.

Research Recognition & Competitions

  • Grand Finalist, QETCI Global Quantum Hackathon

    Top 16 of 1,600+ teams across 25+ countries (quantum ML).

    2022

  • Grand Finalist, Smart India Hackathon (NDRF track)

    Led a team of 6 on VIKAS, a multimodal disaster-response system. 1,000+ teams competed nationally.

    2022

  • 2nd Runner-Up, Cisco Intern Case Study Competition

    NLP supplier-name standardization, 50,000+ records, 100+ entries.

    2024

Scholarships & Honors

  • NTSE State Scholar

    National Talent Search Examination, rank 21 of 151,000+.

    2018

  • Thayil Lonappan George Memorial Endowment Award

    3rd rank, All India Senior School Certificate Examination.

    2020

Selected Schools & Programmes

  • ACM Winter School on Optimization for ML & OR

    Selected participant, IIT Goa.

    2023

  • Summer School on Dynamic Resource Allocation

    Center for Networked Intelligence, IISc Bengaluru.

    2023

Leadership

  • Co-founder & Technical Head, Project Kalpana

    Led a team of 8 building a radio astronomy system, funded by a $13,000 R&D grant.

    2022–23

  • General Secretary & Treasurer, ACM Student Chapter, Manipal

    Organized 12+ events, including workshops and hackathons. Participation grew 35%.

    2022–23

  • Technical Head, Astronomy Club of Manipal

    Led 5+ projects and ran astronomy sessions and stargazing events for 200+ students.

    2021–23

Get in touch

Contact

Email is the best way to reach me about my research, a possible collaboration or anything on this page.

Research methods:
Temporal Graph Neural Networks, Language Models on Graphs, LoRA / Adapters, Representation Learning, Tensor Decomposition, Quantum ML, Transformers / NLP, Computer Vision
Frameworks:
PyTorch, PyTorch Geometric, DGL, DyGLib, TensorFlow, Qiskit, scikit-learn
Systems & tooling:
Python, C++, CUDA, FastAPI, Docker, Kubernetes, Distributed Training
Email
gauthamprabhu9@gmail.com
Google Scholar
Publications & metrics
GitHub
GauthamPrabhuM
LinkedIn
gautham-prabhu
Curriculum Vitae
PDF