Bhavya Minesh Shah

Bhavya Minesh Shah is a Robotics and Artificial Intelligence (AI) Researcher and Courtesy Affiliate at the Logos Lab, Arizona State University (ASU). He completed his Master of Science in Robotics and Autonomous Systems (Artificial Intelligence) in May 2026 through the School of Computing and Augmented Intelligence (SCAI), part of the Ira A. Fulton Schools of Engineering, graduating with a Grade Point Average (GPA) of 3.57 out of 4.0. His work sits at the intersection of agentic AI and robotics, spanning robot learning, computer vision, and multimodal Large Language Models (LLMs).

At the Logos Lab, under Dr. Nakul Gopalan, Bhavya develops and evaluates robot learning policies including Action Chunking Transformers (ACT), diffusion models, and Vision-Language Action (VLA) models trained on teleoperation data from a bi-manual robot. He built an annotation tool with automated annotation capability that uses dynamic finetuning and Reinforcement Learning (RL) of LLMs from annotator feedback. He also benchmarked open-source Automatic Speech Recognition (ASR) models and audio-capable multimodal LLMs for an Emergency Medical Services (EMS) call project, evaluating them on Word Error Rate (WER), Bilingual Evaluation Understudy (BLEU), and Recall-Oriented Understudy for Gisting Evaluation (ROUGE) to determine whether off-the-shelf models required finetuning. His simulation work includes a MuJoCo table tennis environment used to train reinforcement learning policies such as Proximal Policy Optimization (PPO) and Soft Actor-Critic (SAC). He is second author on "Meanings and Measurements: Multi-Agent Probabilistic Grounding for Vision-Language Navigation" (March 2026), available at https://arxiv.org/abs/2603.19166.

From May 2026 to August 2026, Bhavya worked as a Software Engineer in Platform and Integrations at Anton Intelligence LLC in Tempe, Arizona. He migrated local file storage from an Azure Virtual Machine (VM) to Azure Blob Storage, converting it into a Platform as a Service (PaaS) solution that prevented data loss while maintaining compliance. He extended the Azure Structured Query Language (SQL) Database writer in Python to capture metadata for every uploaded file, improving traceability and simplifying audits. He also built a configuration-driven report generation Python package that allowed non-developers to produce reports independently, and deployed a modular forecasting script to Azure Function App through Azure DevOps pipelines to give the platform on-demand predictive capability.

From May 2025 to May 2026, Bhavya was a Helios Scholar and Graduate Intern (AI Researcher) at the Translational Genomics Research Institute (TGen) in Phoenix, Arizona. He designed and developed an agentic AI platform for physicians and biomedical practitioners supporting Health Insurance Portability and Accountability Act (HIPAA) compliant data handling, multimodal reasoning, and Visual Question Answering (VQA) over patient imaging and clinical records. He built an LLM orchestration system using LangChain, LangGraph, and Ollama that coordinated task-specific LLMs and Vision-Language Models (VLMs) for de-identification, diagnosis, recovery tracking, and reasoning segmentation under a single orchestrator model, and developed Reasoning and Referred Segmentation features that make model decision logic visible to clinical users. His research on AI-assisted Tumor Boards involved training multimodal models with Contrastive Language-Image Pre-training (CLIP) encoders on Magnetic Resonance Imaging (MRI) scans, Computed Tomography (CT) scans, X-rays, and patient histories, evaluating MedGemma, MedLlama3, and Qwen3 among others. Physical tumor boards, which convene multiple specialists to review a single case, are costly and infrequent; this line of research aims to shorten that process and extend expert review to far more cases. He optimized large-scale training on High-Performance Computing (HPC) clusters using Simple Linux Utility for Resource Management (SLURM), HuggingFace Transformers, Accelerate, Parameter-Efficient Fine-Tuning (PEFT), and BitsAndBytes, applying Low-Rank Adaptation (LoRA), quantization, and mixed precision for multi-Graphics Processing Unit (GPU) training. He incorporated Attention Pooling, Perceiver Resampling, and a Unified Embedding Decoder using MediaPipe and Open Neural Network Exchange (ONNX) to reduce inference latency, and presented his findings in a scientific poster and oral presentations at the institute's research symposium.

Before beginning his graduate studies, Bhavya was an Artificial Intelligence Engineer at Accurate Industrial Controls Pvt Ltd in Pune, India from August 2023 to May 2024, where he built the company's AI division from scratch. He set up multi-GPU servers, Compute Unified Device Architecture (CUDA) and CuDNN environments, and GPU-compatible Machine Learning (ML) frameworks, establishing the practices the division used for scalable deployment. He developed perception and anomaly detection pipelines, finetuning YOLOv8 and Water Segmentation and Refinement (WaSR) to 96.4% mean Average Precision (mAP) at 0.5 Intersection over Union (IoU) for an autonomous boat program commissioned as a challenge by the Indian Navy, and optimizing YOLOv8, PatchCore, EfficientAD, and EasyOCR with ONNX and TensorRT on the Nvidia Triton Inference Server for real-time defect detection. That defect detection work included a Computer Vision (CV) pipeline for Liquefied Petroleum Gas (LPG) cylinders that determined weight, expiry, and defects, deployed in a Zone 1 hazardous environment where it removed human involvement from the inspection process entirely. He also designed the B-Star path planning algorithm, which improved planning speed by 300% over A-Star, and increased multi-sensor fusion reliability by introducing UNIX timestamp synchronization.

Earlier, Bhavya interned as a Machine Learning Intern at Swasthya.ai, an oncology startup in Mumbai, India, from September 2022 to March 2023. There he worked on Natural Language Processing (NLP), including LLMs and attention mechanisms, extracting clinical entities from medical notes and using them to train models such as Bidirectional Encoder Representations from Transformers (BERT).

Bhavya received his Bachelor of Technology in Computer Science and Engineering with a Bioinformatics specialization from Vellore Institute of Technology (VIT), Vellore, India, in 2023. His coursework and projects spanned Artificial Intelligence, Machine Learning, Deep Learning, Natural Language Processing, Image Processing, Web Development, Operating Systems, and Networking. He published "Demonstrated Deep Learning Techniques for the Resolution of CAPTCHA Images" in the International Research Journal of Engineering and Technology (IRJET) in September 2022, and for his capstone project introduced a novel object detection technique built on semantic segmentation masks.

During his graduate studies, Bhavya served as Technical Officer of ASU's AI Society, where he delivered a lecture on Semantic Segmentation, available at https://www.youtube.com/watch?v=3O6Be4hoLMM, and a second lecture on 3D Computer Vision.

Outside of academic and professional work, Bhavya trekked to the Brahmatal summit at over 12,800 feet in January 2022 and the Sandakphu and Falut summits at over 11,800 feet in January 2023. He swims and hikes regularly and follows cricket, table tennis, badminton, and snooker. He played cricket at the district level and was a member of his school team.