Harshitha Manjunatha

PhD Researcher @ UT Dallas | AI/ML Engineer | Multimodal LLMs & UAV/Edge Intelligence

My research focuses on building data-efficient perception systems for real-world environments, with particular emphasis on active learning in LLMs, amodal reasoning, and UAV-based intelligence. I work at the intersection of vision-language models, hyperspectral imaging, and edge-cloud deployment to develop scalable AI solutions for precision agriculture and autonomous systems. Prior to my PhD, I worked as a Data Scientist at BMW Group and an AI Consultant at Deloitte Germany experiences that continue to shape my approach to translating foundational AI research into practical, deployable systems.

Achievements

During my first semester in my current PhD program, I successfully secured a High-Performance Computing (HPC) grant of 750,000 ACCESS credits at the Texas Advanced Computing Center (TACC). I am currently leveraging this grant to train custom object detection models that can be used on edge devices with resource constraints.

Education

PhD in Computer Engineering - University of Texas at Dallas (2025 - Current)
MS in Information & Communication Systems - Technical University of Chemnitz, Germany
BS in Electronics & Communication Systems - Visvesvaraya Technological University, India

Campus Partnerships: Perplexity 2025 | Microsoft Student Ambassador 2025

Learn more about my:

Skills

Technical expertise & tools

Experience

Professional journey

Publications

Research papers

Conferences

Talks & poster sessions

Projects

Portfolio highlights

Conferences

Research dissemination, poster sessions, and professional engagement

Poster presentation at IPPS Southern Region (2025)

IPPS Southern Region conference group photo (2025)

Poster Presentation Oct 2025

IPPS Southern Region of North America — Charles Student Research Competition

Listed as a poster presenter in the official IPPS Southern Region announcement (28 Oct 2025).

Poster

AI-Driven Tool for High-Dimensional Phenotyping of Floral Health in Roses Using Hyperspectral Imaging

University of Texas at Dallas

Presented a research poster on hyperspectral imaging + AI for floral health phenotyping.
Engaged with horticulture and nursery industry researchers to discuss applied AI for crop production.

View official announcement

Research Day · Poster Session Nov 2025

UT Dallas — ECE Research Day

Shared ongoing PhD research at the Erik Jonsson School of Engineering and Computer Science, presenting work from the Center for Robotics and Intelligent Systems (CeRIS).

Focus

Multimodal vision-language systems for precision agriculture (UAV + ground robots)

The University of Texas at Dallas · ECE

Built multimodal drone/robot + LLM/VLM pipelines for plant monitoring and decision support in large-scale nursery operations.
Adapted YOLO-family detectors for hyperspectral UAV imagery (281 bands), prioritizing efficient edge/on-device inference and domain adaptation under real field variability.
Emphasized multimodality beyond text—integrating UAV and ground-robot sensing with vision-language reasoning to align model outputs with operational context in the field.
Leveraged HPC workflows (e.g., Slurm/TACC-style scheduling) to scale training and evaluation across large hyperspectral datasets while maintaining reproducibility and throughput.
Grateful to Lakshman Tamil and Prabha Sundaravadivel, Ph.D., for mentorship and guidance.

View event page

Presenting at the International Conference on Large Language Models (LLM 2025), Las Vegas

Oral Presentation Dec 2025 · Las Vegas, USA

International Conference on Large Language Models (LLM 2025)

Presented a talk on multimodal agricultural intelligence representing The University of Texas at Dallas and the Center for Robotics and Intelligent Systems (CeRIS).

Talk

Cutting Through the LLM Noise: Hardware-aware multimodal AI for precision agriculture

UT Dallas · CeRIS

Argued for next-wave LLM systems that are context-aware and grounded in real-world constraints (latency, memory, and deployment substrate).
Positioned the research frontier at the intersection of LLMs × hardware × operational optimization for field robotics and aerial sensing.
Grateful to Lakshman Tamil and Prabha Sundaravadivel, Ph.D., for mentorship and support.

View conference page

Microsoft Ambassador Program activity at UT Dallas Office of Information Technology

Perplexity Campus Partner activity at UT Dallas

Campus Partnerships 2025

Campus Partner — Perplexity (2025) · Microsoft Ambassador — UT Dallas OIT

Selected for two student-facing programs focused on translating emerging AI capabilities into practical, community-centered impact on campus.

Roles

Perplexity Campus Partner (Campus Strategist Program) · Microsoft Ambassador Program (UT Dallas Office of Information Technology)

Leadership · Community building · Applied AI literacy

Designed and supported on-campus engagements to drive responsible, high-signal adoption of AI tools and workflows.
Worked across student cohorts and campus stakeholders to communicate practical usage, constraints, and best practices.

UT Dallas OIT program Perplexity program Microsoft Learn program

Publications

Research papers and conference proceedings

TePD: Temporal Privileged Distillation for Amodal Counting Under Structured Occlusion

H. Manjunatha, P. Sundaravadivel, S. Borah, H. Torbert, L. Tamil

In Proceedings of ICML, 2026. Submitted

Beyond Green: Domain-Aware Self-Validated Instance Counting for Loropetalum and Non-Green Ornamental Species using Annotation-Free Robustness Metrics

H. Manjunatha, P. Sundaravadivel, S. Borah, L. Tamil, P. Knight, H. Torbert, S. Kumpatla

In Proceedings of CVPR, Jun 2026. Submitted

Comparative Analysis of Large Language Models for Multimodal Agricultural Intelligence in Precision Nursery Management Systems

H. Manjunatha, S. Borah, A. Anand, P. Sundaravadivel, L. Tamil, S. Kumpatla, P. Knight, H. Torbert

In Proceedings of LLM Conference, Dec 2025. Accepted

Ros-AI: An LLM-Enhanced Scalable Multimodal Framework for UAV-Based Rose Bloom Analysis in Precision Agriculture

P. Sundaravadivel, H. Manjunatha, K. Narasimhamurthy, et al.

TechRxiv Preprint, Oct 2025. Preprint DOI: 10.36227/techrxiv.176108108.86171044/v1

MagCount: A Digital Framework for Automated Counting of Southern Magnolias for Inventory Management in Nurseries

M. Syed, A. Anand, S. Borah, H. Manjunatha, P. Sundaravadivel, P. Knight, P. Gowda, H. Torbert

In Proceedings of IPPS, Oct 2025. Accepted

AI-driven Tool for High-dimensional Phenotyping of Rose Flowers using Hyperspectral Imaging

H. Manjunatha, P. Sundaravadivel, S. Borah, L. Tamil, S. Kumpatla, P. Knight, H. Torbert

In Proceedings of IPPS, Oct 2025. Accepted

Projects

Highlighting key research and development projects

Annotation-Free Tomato Disease Detection

Novel diagnostic framework leveraging unsupervised computer vision and LLMs for automated disease detection.

UAV-Based Rose Bloom Analysis

LLM-enhanced multimodal framework achieving 4.7x improvement in detection precision.

Instance Counting for Ornamental Species

Domain-aware self-validated counting system for non-green plants using annotation-free metrics.

SPC Digital Assistant for SAP TechOps

Azure AI and GPT-4 powered assistant achieving 30% productivity increase.

Automotive ITSM Chatbot

State-of-the-art chatbot using AWS Bedrock and RAG for automotive service management.

EV Battery Health Prediction

LSTM-based deep learning system for predicting battery patterns and end-of-life indicators.

Hyperspectral Imaging for Phenotyping

High-dimensional analysis tool for rose flower phenotyping using hyperspectral imaging.

Edge AI Deployment on Jetson Nano

Hybrid edge-cloud architecture for real-time inference on NVIDIA Jetson Nano platforms.

MagCount: Magnolia Counting

Digital framework for automated counting of Southern Magnolias for inventory management.

To see more of my projects...

Visit My GitHub

Skills

Technical skills and certifications

Technical Skills

AI/ML

Large Language Models (LLMs) Generative AI RAG Computer Vision Deep Learning LSTM TinyML Edge AI

Programming

Python C++ JavaScript TensorFlow PyTorch YOLO scikit-learn OpenCV ROS Pandas NumPy

Cloud & HPC

AWS (Bedrock, SageMaker) Azure (OpenAI) GCP (Vertex AI) TACC (Stampede3, Vista) Slurm

Deployment

NVIDIA Jetson Nano ESP32 UAV/Drone Systems Ground Robots Edge Devices GPU Computing Real-Time Inference

Tools

Docker Kubernetes Git Databricks NVIDIA Nemo Guardrails MLflow Hugging Face Unix/Linux

AI-Native Dev Tools

Cursor Replit Lovable Bolt.new Emergent n8n

Agentic AI & Protocols

OpenAI Swarm CrewAI A2A MCP DSPy GraphRAG Multi-Agent Systems

Certifications

AWSCloud Practitioner
AWSMachine Learning Engineer
NVIDIANinja Tech Strategist
NVIDIALLMs Specialist
Google CloudGenAI
Google CloudVertex AI Certification
DatabricksAI Specialist
IBMData Scientist
MicrosoftInfrastructure Engineer
GovernanceAI Governance Bootcamp

Experience

Graduate Research Assistant | PhD Researcher

University of Texas at Dallas, Dallas, TX

Aug. 2025 – Present

Formulated novel annotation-free diagnostic framework for tomato pathology leveraging unsupervised computer vision techniques with large language models for automated disease detection and classification
Built hybrid segmentation module isolating plant subjects from complex backgrounds by fusing Laplacian-based focus analysis with colorimetric clustering in HSV color space (Google Lens concept)
Developed multi-modal diagnostic model quantifying pathological anomalies via CIELAB and Local Binary Pattern (LBP) analysis, leveraging locally-hosted LLM Ollama to generate actionable agronomic recommendations
Conducted rigorous comparative analysis of object detection architectures, systematically testing models from YOLOv5 to YOLOv11 and SOD-YOLO on Roboflow to benchmark efficacy for small object detection in aerial UAV imagery
Engineered 4.7x improvement in mean Average Precision for rose detection by creating custom training on UAV data with strategically annotated dataset, amplified through extensive data augmentation routines
Integrated and benchmarked multiple large-scale vision-language models (Google Gemini, Meta Llama 3.2) to construct automated agricultural intelligence system for detailed crop management reports
Investigated hybrid edge-cloud deployment practicalities, transitioning from cloud-based services to local infrastructure for training while exploring on-device inference on NVIDIA Jetson Nano platforms

AI Consultant

Deloitte Germany, Munich, Germany

Sep. 2022 – Apr. 2025

Coordinated the development and implementation of data-driven used car strategy prototype, improving procedures to grow revenues and shift management perspectives
Strategized and led GenAI Innovation Labs, driving successful AI transformations for major automotive clients through data-driven oversight
Designed and implemented SPC Digital Assistant for SAP TechOps using Azure AI and GPT-4, streamlining internal processes by delivering precise answers from SAP wiki content and HANA database tickets, achieving 30% productivity increase and 20% faster ticket resolution; integrated with SAP's Joule AI copilot
Engineered state-of-the-art Automotive ITSM chatbot using AWS Bedrock services and RAG functionality
Integrated NVIDIA's Nemo Guardrails technology into GenAI services on NVIDIA GPU Cloud registry, resulting in enhanced interactive vehicle manuals and increased customer satisfaction
Strategized with automotive clients to revise cost optimization strategies for Software Defined Vehicles (SDVs) leveraging GenAI, saving €2M in data platform costs
Implemented validation and evaluation pipeline for LLMs with pricing optimization, ensuring quality results
Developed GenAI blueprints for cloud-based (AWS & GCP) PoCs, facilitating rapid deployment

Data Scientist

BMW Group, Munich, Germany

Apr. 2020 – Aug. 2022

Actively contributed to lifecycle administration by extracting critical data about electric batteries, incorporating assessment steps to assure accuracy and quality of data collected
Researched extensively into battery state of health and executed root cause analysis for faults and failures in battery lifecycle, instrumental in understanding the battery cycle of EVs
Built and integrated clustering machine learning model for customer segmentation and assessing State of Health (SOH) of batteries
Applied advanced deep learning techniques like Long Short-Term Memory Networks (LSTMs) to predict battery patterns in different scenarios and analyze end-of-life indicators
Implemented cloud-based data management for automakers, optimizing multinational team collaboration and streamlining real-world and simulated test drive data handling

Contact Me

Email:

harshitha.manjunatha@outlook.com

Harshitha Manjunatha

PhD Researcher @ UT Dallas AI/ML Engineer Multimodal LLMs & UAV/Edge Intelligence