SBSumit Barua Let’s connect ↗
Professional portrait of Sumit Barua
AI Researcher Kalamazoo, Michigan

Graduate Research Assistant - Computer Science · Western Michigan University

Sumit Barua

I build trustworthy AI systems that explain their decisions, communicate uncertainty, and remain grounded in reliable evidence.

Seeking Fall 2027 Ph.D. opportunities. I am interested in trustworthy AI, explainable AI, multimodal learning, and high-stakes AI applications.

Google Scholar Curriculum Vitae GitHub LinkedIn Email
Trustworthy AI AI Explainability Multimodal AI Evidence-Grounded LLM AI for Healthcare
2 Published first-author
journal articles
5 Manuscripts under review
6+ End-to-end AI systems
built and evaluated
1
Research Agenda
Trustworthy AI under uncertainty and domain shift

01 / Research agenda

Trustworthy AI from
data to decision.

My research asks a central question: how can AI systems recognize unreliable inputs, explain their reasoning, and justify their outputs before people act on them?

01 Reliable input
02 Transparent model
03 Verified decision
Layer 01

Input reliability

Data reliability and
distribution shift

Can an AI system recognize when its evidence is incomplete, shifted, or unreliable?

I develop methods to measure documentation quality, identify behavioral changes, and determine whether incoming data are reliable enough to support automated decisions.

Methods

  • Reliability scoring
  • Distribution shift
  • Behavioral analysis
  • Data-quality assessment
Representative work

EMS-RRI · IoMT intrusion detection

Layer 02

Model transparency

Explainable and
uncertainty-aware AI

Can a model explain what influenced its prediction and communicate when that prediction should be questioned?

I combine explainability, uncertainty awareness, and multimodal learning to make model behavior more interpretable in high-stakes environments.

Methods

  • Explainable AI
  • Uncertainty estimation
  • Multimodal learning
  • Grad-CAM
  • LLM-XAI
Representative work

Wound assessment · Smart-ambulance AI

Layer 03

Output verification

Evidence-grounded
decision support

Can every AI-generated recommendation be traced to valid evidence before it influences a decision?

I build language systems that validate citations, detect unsupported claims, regenerate unreliable responses, and request human clarification when sufficient evidence is unavailable.

Methods

  • Hybrid retrieval
  • Citation validation
  • Bounded regeneration
  • Human oversight
Representative work

CiteGuard-RAG · SpaceRAG VR

Unifying goal

Across healthcare, cybersecurity, and language AI, my research moves beyond predictive accuracy to examine whether an AI system’s data, reasoning, and evidence are reliable enough to support real-world decisions.

02 / Selected publications

Research in the field.

2026 Cureus Journal Published

Uncertainty-Aware Wound Localization and First-Aid Decision Support From Smartphone Images Using an Explainable Multimodal AI Framework

An explainable, human-in-the-loop framework integrating YOLOv11 wound segmentation, ResNet-50 body-location classification, and Mistral-7B first-aid generation, with Grad-CAM and LLM-XAI explanations and clarification requests for unrecognized visual outputs.

Barua, S., Bahre, R., Babiceanu, R., & Hong, G. (2026)
Computer Vision · Large Language Models
2026 Discover Public Health Under review

Quantifying Behavioral Contributions to Diabetes Risk Using an Explainable Multiclass Machine Learning Framework with BRFSS 2024

This study develops an explainable multiclass framework for BRFSS 2024 diabetes-risk analysis, showing how behavioral/health-status factors contribute to model-attributed diabetes risk while exposing the difficulty of identifying prediabetes from self-reported surveillance data.

Barua, S., Jabran, S., Rodgers, C. & Fong, A. (2026)
Explainable AI · Public Health Surveillance · Diabetes Risk Modeling
2026IEEE Communications MagazineUnder review

From Accuracy to Message Trust: Operational Validation of Quantum-AI Pipelines for V2X Communication Security

This research designs and validates a trustworthy Quantum-AI pipeline for V2X communication security, showing that reliable message-trust decisions require external transfer testing, unseen-attack evaluation, uncertainty-aware routing, and failure analysis beyond internal accuracy.

Barua, S., Fong, A., Fong, B., & Kim, H. (2026)
QML · V2X Communications · Message Security
2026 AI Under review

Applicability-Aware Reliability Scoring and Prediction of EMS Patient-Care Records for Trustworthy Smart Ambulance AI

This research introduces the EMS Record Reliability Index (EMS-RRI), an interpretable reliability-screening framework for smart-ambulance AI, showing that trustworthy prehospital automation requires evaluating whether EMS patient-care records are complete enough before downstream triage, routing, resource coordination, or clinical decision-support models are used.

Barua, S., & Mahmud, B.U. (2026)
Trustworthy AI · EMS Data Quality · Smart Ambulance
2026Wiley Security and PrivacyUnder review

Behavior-Aware Explainable AI for IoMT Intrusion Detection Under Distribution Shift

Interpretable network-behavior modeling designed to reveal how intrusion detectors generalize beyond familiar attacks and datasets.

Barua, S.Cybersecurity · XAI · Dataset Shift
2026 arXiv Preprint

Commonsense Reasoning in Computer Vision: Foundations, Recent Advancements, and Future Directions

This paper presents a comprehensive survey of recent developments that integrate commonsense knowledge into computer vision tasks. We systematically review approaches based on knowledge graphs, scene graphs, neuro-symbolic models, and commonsense-augmented transformers.

Mahmud, B. U., Barua, S., Hong, G. Y., Gupta, A., & Liu, H.
Computer Vision · Commonsense Reasoning · Vision Language Models

Updates

News

Project

Completed the first version of SpaceRAG VR, a Meta Quest learning environment for teaching evidence retrieval, reranking, citation validation, and responsible reliance on AI-generated answers.

Paper

Submitted our paper From Accuracy to Message Trust: Operational Validation of Quantum-AI Pipelines for V2X Communication Security to IEEE Communications Magazine

Paper

Submitted CiteGuard-RAG: A Validation-Centered AI System for Evidence-Grounded Question Answering to Engineering Reports.

Paper

Our manuscript on Behavior-aware Explainable AI for IoMT Intrusion Detection Under Distribution Shift entered peer review at Security and Privacy.

04 / Selected projects

Research translated into
working systems.

Selected end-to-end systems through which I investigate trustworthy, explainable, and evidence-grounded AI in high-stakes applications.

CiteGuard-RAG evidence-grounded question-answering pipeline 01 Evidence-grounded language AI
Research system Retrieval-augmented generation

CiteGuard-RAG

A validation-centered question-answering system designed to produce evidence-grounded responses while detecting unsupported claims and unreliable citations.

Research contribution

I designed and evaluated the complete pipeline, including hybrid retrieval, evidence reranking, sentence-level citation validation, bounded regeneration, and external domain testing.

  • Hybrid retrieval
  • Evidence reranking
  • Citation validation
  • Bounded regeneration
  • Domain-shift evaluation
98.3% Citation validity in controlled evaluation
3 Evaluation domains
View on GitHub
SpaceRAG VR trustworthy-AI learning environment 02 Immersive trustworthy-AI education
Active project Unity and virtual reality
National Science Foundation CyberTraining Project Award #2320951 Official award

SpaceRAG VR

An immersive Meta Quest learning environment where students experience evidence retrieval, reranking, answer validation, and the consequences of relying on unsupported AI-generated information.

My contribution

I contribute to the project through learning-experience design, Unity development, trustworthy-AI workflow integration, gameplay validation, and Meta Quest deployment.

  • Unity 6
  • Meta Quest
  • OpenXR
  • Trustworthy-AI education
  • Gameplay validation
View on GitHub
Explainable multimodal AI system for wound assessment 03 Explainable healthcare AI
Published research Multimodal decision support

Explainable Multimodal Wound AI

A human-in-the-loop decision-support system that connects wound segmentation and body-location classification with explainable first-aid generation.

Research contribution

I integrated YOLOv11 wound segmentation, ResNet-50 body-location classification, and Mistral-7B first-aid generation with Grad-CAM, LLM-XAI, and clarification requests for unrecognized visual outputs.

  • YOLOv11
  • ResNet-50
  • Mistral-7B
  • Grad-CAM
  • LLM-XAI
  • Human in the loop
View on GitHub View publication

05 / Contact

Let’s build AI
that earns trust.

I am welcoming conversations about research collaboration and applied work in trustworthy, explainable, and evidence-grounded AI.

Best way to reach me sumit dot barua at wmich dot edu

Open to conversations about

  • 01 Fall 2027 Ph.D. opportunities
  • 02 Trustworthy-AI research collaborations
  • 03 Multimodal and evidence-grounded AI