Arjun Vaghasiya

Spam and Abuse Intelligence Team

• Adapted Microsoft’s MAGMA vision-language foundation model for robot manipulation tasks in industrial manufacturing, fine-tuning the VLA on NVIDIA H100 GPU to improve domain-specific accuracy and robustness while leveraging spatial and temporal reasoning for robotic action planning. • Developed and validated a robotic perception-to-action pipeline from camera input through MAGMA inference to decision-making in NVIDIA Isaac Sim, enabling scalable simulation and iterative testing prior to deployment. • Engineered and optimized robot–camera configurations (end-effector, RGB/RGB-D/stereo setup, placement strategies) to enhance perception reliability in complex assembly scenarios. • Constructed a domain adaptation dataset combining semi-synthetic and real-world assembly images with ground-truth annotations, enabling systematic evaluation of inference accuracy, latency, and error recovery. • Deployed the MAGMA-powered pipeline on a FANUC industrial robot, achieved 52.3% zero-shot manipulation accuracy and 64.1% spatial reasoning performance, aligned with MAGMA’s benchmark results.

• Engineered an end-to-end, clinically oriented keypoint detection framework for foot radiographic analysis, leveraging a customized YOLO-based architecture with advanced feature fusion to achieve highly precise anatomical landmark localization under low-contrast and high-noise imaging conditions. • Developed a multi-scale hybrid deep learning architecture integrating Swin Transformer and ResNet-50, enabling joint modeling of long-range contextual dependencies and fine-grained spatial features, resulting in 91% diagnostic classification accuracy across complex musculoskeletal pathologies. • Designed and deployed an automated clinical narrative generation pipeline using GPT-2, translating multimodal model outputs into structured, human-readable medical reports while ensuring alignment with radiological reporting standards. • Incorporated large language model–driven semantic refinement and consistency validation, enhancing diagnostic expressiveness, reducing ambiguity, and improving the clinical actionability of AI-generated medical documentation within real-world healthcare workflows.

• Built and deployed an AI-powered Intelligent Document Processing (IDP) platform with a RAG pipeline and Vector DB, reducing manual review effort by 50% (40% to 20%) through adaptive workflows and automation. • Developed and productionized multi-step AI agents with LangChain and LangGraph, integrating a RAG pipeline to enable real-time retrieval, reasoning, streaming, and human-in-the-loop document analysis. • Ensured reliability and compliance of LLM applications using LangSmith for monitoring, evaluation, and observability, while deploying LangGraph apps as APIs and intelligent assistants for enterprise automation. • Applied advanced LLM optimization techniques (prompt engineering, RLHF, DPO, LoRA/PEFT/QLoRA) and fine-tuned the Llama 3 model, achieving a 20% performance gain while lowering training costs. • Optimized deep learning models for enterprise-scale workloads, reducing inference latency by 30% with distillation, pruning, and quantization to accelerate high-volume document processing.

• Led end-to-end development of an AI-powered customer engagement platform, integrating recommendations, semantic search, TTS assistants, and computer vision to enhance multi-channel user experiences. • Built and deployed large-scale recommendation systems (collaborative filtering, deep ranking models, SQL-driven data pipelines) serving thousands of users daily, increasing CTR by 15% and boosting order value by 12%. • Developed semantic search using Sentence-BERT, Elasticsearch and voice-based assistants (Tacotron 2, WaveGlow, ASR models), reducing bounce rate by 20%. • Delivered computer vision models (YOLOv5, ResNet) for automated product tagging and quality inspection with 95%+ precision across millions of catalog images. • Optimized and scaled ML pipelines with ONNX, pruning, quantization, MLflow, Airflow, Docker, Kubernetes, AWS, cutting inference latency by 30%, compute costs by 20%, and scaling to 100k+ monthly API requests with 99.9% uptime.

Demonstrated outstanding academic performance with a perfect 4.0 GPA