PhD Candidate · University of Rochester

Akhil
Kasturi

Akhil Kasturi

Building large language models, multimodal deep learning systems, and agentic AI workflows. My research focuses on LLM pre-training and domain adaptation, cross-modal reasoning across text, vision, and time-series data, and multi-agent orchestration for automated generation tasks.

Research Interests: LLM Pre-Training & Fine-Tuning, Agentic AI, Multimodal Fusion, Self-Supervised Learning, NLP, Vision-Language Models

01 — About

Building intelligent
systems at scale

I am a PhD candidate at the University of Rochester developing large language models, multimodal AI systems, and agentic workflows for high-stakes applications.

My research spans LLM pre-training and domain adaptation on billion-scale corpora, multimodal fusion systems that jointly reason across text, vision, and time-series signals, and multi-agent LLM orchestration for automated report generation. I work across the full stack—from distributed GPU training on AWS to evaluation and deployment.

I have published 22 peer-reviewed papers (10+ first-author), received the Best Scientific Poster Award at SPIE Medical Imaging, and currently have a manuscript under review at Nature Scientific Reports. I am advised by Prof. Axel Wismüller.

Education

PhD, AI for Multimodal Learning & Generative AI (Dept. of ECE)
University of Rochester (In Progress)

M.S., AI and Machine Learning (Dept. of ECE)
University of Rochester

M.Eng., Advanced Signal Processing (Dept. of ECE)
Western University

Core Skills

PyTorch · Hugging Face Transformers · DeepSpeed · TensorFlow · Python · C++ · AWS (EC2, S3) · Docker · CUDA · Multi-GPU Distributed Training (A100, H100)

Contact

akasturi@ur.rochester.edu

02 — News

Latest updates

February 2026

SPIE Medical Imaging 2026 — LLM Fine-Tuning for Structured Data Classification

Presented poster on fine-tuned LLaMA-based large language models trained on structured tabular variables for rapid, actionable risk stratification from patient records.

February 2026

SPIE Medical Imaging 2026 — Causal Feature Attribution for LLMs (Oral)

Delivered oral presentation on a causal feature ablation methodology to quantify how individual input variables influence fine-tuned LLM predictions across structured and unstructured text data.

December 2025

RSNA 2025 — Multimodal Reasoning Presentation

Presented research on multimodal fusion methods combining LLM-encoded text, 3D Vision Transformers, and time-series signal representations at the Radiological Society of North America Annual Meeting in Chicago.

November 2025

Multi-Institutional Collaboration — Billion-Scale Dataset

Partnering with Yale University and UC Irvine under NIH funding to build one of the largest multimodal datasets in the domain—integrating imaging, time-series signals, and 4 billion+ text records across 3 independent sites.

03 — Research

Selected projects

02

Multimodal Cross-Modal Reasoning

Built a novel multimodal reasoning model fusing LLM text embeddings, graph-based time-series encoders, and 3D Vision Transformer video representations via cross-modal attention. Achieved 97% AUC with cross-site generalization on 2,330 samples from 3 independent institutions.

PyTorch · 3D-ViT · GNN · Contrastive Learning
03

Agentic AI for Automated Report Generation

Designed a multi-agent orchestration system using Gemini-based LLM agents for end-to-end automated report generation. Implemented agent communication protocols, tool-use integration, and RAG with domain knowledge bases. Evaluated via BERTScore, ROUGE, and human-expert review.

Gemini Agents · RAG · BERTScore · Docker
04

Causal Feature Attribution for LLMs

Developed a causal feature ablation methodology to quantify how individual input variables influence fine-tuned LLM predictions, systematically isolating each feature's contribution across structured tabular and unstructured text data for interpretable, trustworthy AI.

SPIE 2026 Oral · Explainable AI
05

Self-Supervised Learning & Foundation Model Adaptation

Applied DINO self-distillation to pre-train Vision Transformers on unlabeled image datasets, reducing labeled data needs by 80–90%. Fine-tuned SAM for instance segmentation with DINO-initialized weights. Shared backbone for detection and segmentation tasks.

DINO · SAM · ViT · PyTorch
06

Vision-Language Report Generation

Built BioVLM-T, a generative vision-language foundation model incorporating temporal prior images and patient history for automated report generation. Trained on a large-scale dataset using PyTorch and Hugging Face Transformers.

SPIE Medical Imaging 2025
07

Transformer-Based Keypoint Detection

Developed transformer-based methods for precise keypoint localization in images using heatmap regression heads. Leveraged DINO self-supervised pre-trained features for weight initialization. Won Best Scientific Poster Award at SPIE Medical Imaging 2024.

Best Paper Award · SPIE 2024
04 — Conferences

Conference gallery

04B — Posters

Poster presentations

SPIE Medical Imaging — Posters

Poster: Fine-Tuned LLMs for Structured Data Classification
Poster · SPIE Medical Imaging 2026
Fine-Tuned LLMs for Structured Data Classification
Fine-tuned LLaMA-based models on structured tabular variables to enable rapid, actionable classification without unstructured text latency.
Poster: Transformer-Based Keypoint Detection
Poster · SPIE Medical Imaging 2024
Transformer-Based Keypoint Detection Using Vision Transformers
TransUNet-based keypoint localization with Gaussian heatmaps; robust detection under noisy and complex imaging conditions.
05 — Publications

Research output

2025

Multimodal Contrastive Prognostication Framework for Early Neurological Outcome Prediction in Post-Cardiac Arrest Patients (CLAIR)

Akhil Kasturi, Ashley R. Proctor, Ali Vosoughi, et al.

Under Review — Nature Scientific Reports

Preprint →

Masked autoencoders for early neurological outcome prediction in post-cardiac arrest patients using brain CT scan

Akhil Kasturi, Ashley R. Proctor, Ali Vosoughi, et al.

Emerging Topics in Artificial Intelligence (ETAI) 2025

Paper →

BioVLM-T: A temporal framework for radiology report generation using pre-trained vision language foundational models

Akhil Kasturi, Ali Vosoughi, Nathan Hadjiyski, Axel Wismüller

SPIE Medical Imaging: Clinical and Biomedical Imaging 2025

Paper →

ETT-LDx: Transformer-based landmark detection system for endotracheal tube placement verification in chest radiographs

Akhil Kasturi, Ali Vosoughi, Nathan Hadjiyski, Axel Wismüller

SPIE Medical Imaging: Computer-Aided Diagnosis 2025

Paper →

Inferring causal relations from multivariate data using Large-Scale Augmented Granger Causality (lsAGC)

Axel Wismüller, Ali Vosoughi, Akhil Kasturi

NeuroImage (Elsevier), 2025

Paper →

Uncertainty quantification and out-of-distribution detection in skin and breast lesion diagnostics using conformal prediction

Nathan Hadjiyski, C. Kanan, Ali Vosoughi, Akhil Kasturi, Axel Wismüller

Emerging Topics in Artificial Intelligence (ETAI) 2025

Paper →

Large-scale nonlinear Granger causality (lsNGC) analysis of functional MRI data for schizophrenia classification

Axel Wismüller, Ali Vosoughi, Akhil Kasturi

SPIE Medical Imaging: Computer-Aided Diagnosis 2025

Paper →

Analysis of brain connectivity in autism spectrum disorder using large-scale non-linear Granger causality (lsNGC)

Axel Wismüller, Ali Vosoughi, Akhil Kasturi

SPIE Medical Imaging: Clinical and Biomedical Imaging 2025

Paper →

Large-scale augmented Granger causality (lsAGC) for enhanced analysis of brain connectivity in autism spectrum disorder

Axel Wismüller, Ali Vosoughi, Akhil Kasturi

SPIE Medical Imaging: Clinical and Biomedical Imaging 2025

Paper →

2024

Functional connectivity-based classification of autism spectrum disorder using mutual connectivity analysis with local models

Akhil Kasturi, Ali Vosoughi, Nathan Hadjiyski, Larry Stockmaster, Axel Wismüller

Emerging Topics in Artificial Intelligence (ETAI) 2024

Paper →

Anatomical landmark detection in chest x-ray images using transformer-based networks

Akhil Kasturi, Ali Vosoughi, Nathan Hadjiyski, Larry Stockmaster, William J. Sehnert, Axel Wismüller

SPIE Medical Imaging: Computer-Aided Diagnosis 2024

Paper →

Classification of endotracheal tube position in chest x-rays images

Akhil Kasturi, Ali Vosoughi, Nathan Hadjiyski, Larry Stockmaster, William J. Sehnert, Axel Wismüller

SPIE Medical Imaging: Clinical and Biomedical Imaging 2024

Paper →

Segmentation of catheter tubes and lines in chest x-rays using deep learning models

Akhil Kasturi, Ali Vosoughi, Nathan Hadjiyski, Larry Stockmaster, William J. Sehnert, Axel Wismüller

SPIE Medical Imaging: Clinical and Biomedical Imaging 2024

Paper →

Leveraging a memory-driven transformer for efficient radiology report generation from chest x-rays to establish a quantitative metric

Nathan Hadjiyski, Akhil Kasturi, Ali Vosoughi, Axel Wismüller

Emerging Topics in Artificial Intelligence (ETAI) 2024

Paper →

Enhancing graph attention neural network performance for marijuana consumption classification through lsAGC analysis of functional MR images

Ali Vosoughi, Akhil Kasturi, Axel Wismüller

SPIE Medical Imaging: Clinical and Biomedical Imaging 2024

Paper →

Graph attention transformers and large-scale Granger causality to classify marijuana consumption from functional MR images

Ali Vosoughi, Akhil Kasturi, Axel Wismüller

SPIE Medical Imaging: Clinical and Biomedical Imaging 2024

Paper →

2023

Detecting landmarks in anatomical medical images using transformer-based networks

Akhil Kasturi, Ali Vosoughi, Nathan Hadjiyski, Larry Stockmaster, William J. Sehnert, Axel Wismüller

Emerging Topics in Artificial Intelligence (ETAI) 2023

Paper →

Leveraging large-scale Granger causality and neural networks to measure the level of consciousness in DOC patients

Ali Vosoughi, T. Raiser, T. Luther, et al., Akhil Kasturi

Emerging Topics in Artificial Intelligence (ETAI) 2023

Paper →

Identification of schizophrenia patients using large-scale extended Granger causality (lsXGC) in functional MR imaging

Axel Wismüller, Ali Vosoughi, Akhil Kasturi, Nathan Hadjiyski

SPIE Medical Imaging 2023

Paper →

Large-scale Granger causality (lsGC) for classification of schizophrenia using functional MRI

Axel Wismüller, Ali Vosoughi, Akhil Kasturi, Nathan Hadjiyski

SPIE Medical Imaging 2023

Paper →

Large-scale augmented Granger causality (lsAGC) for discovery of causal brain connectivity networks in schizophrenia patients using functional MRI neuroimaging

Axel Wismüller, Ali Vosoughi, Akhil Kasturi, Nathan Hadjiyski

SPIE Medical Imaging 2023

Paper →

Classification of schizophrenia using large-scale kernelized Granger causality (lsKGC) and functional MR imaging

Ali Vosoughi, Akhil Kasturi, Nathan Hadjiyski, Axel Wismüller

SPIE Medical Imaging: Computer-Aided Diagnosis 2023

Paper →

Tracking the impact of global iodinated contrast agent shortage on radiology: analysis of CT exam volumes at a major US healthcare system

Axel Wismüller, J. Avondo, Larry Stockmaster, Akhil Kasturi, Ali Vosoughi

SPIE Medical Imaging 2023

Paper →