Open to opportunities

Utkarsh Dwivedi

@utkarshdwivedi1

AI Engineer and Research Scientist building multimodal, agentic systems and large-scale LLM/MLLM applications from research to deployment.

United States

Message

What I'm looking for

I’m looking for a team where I can build agentic, multimodal AI products end-to-end—designing models, RAG and evals, and shipping scalable services that are reliable, human-centered, and measurable.

I’m an AI Engineer and Research Scientist with 10 years’ experience building multimodal and agentic systems across Computer Vision, NLP, and Generative AI. I focus on architecting and deploying large-scale LLMs and MLLMs for education, legal compliance, and automated Agentic applications.

I’ve developed and deployed generative AI pipelines using Python, PyTorch, Hugging Face, LangChain, and LLaMA models for encoder-only, text generation, and multimodal systems. I work end-to-end across exploratory analysis, data curation, feature engineering, model training, validation, and deployment—backed by a track record of 200+ citations (Neurips & CHI), 5 patents, and multiple fellowships and honorable mentions.

I bring software engineering rigor to AI product delivery, applying Test Driven Development, Refactoring, SOLID, Object Oriented Programming, and Microservices to build scalable, maintainable systems. I also build web-based interfaces and agentic dashboards using React, Node, GraphQL, TypeScript, and Y.js to support collaborative AI applications.

My core technical strengths include LLM fine tuning, retrieval augmented generation (RAG), embeddings, re ranking, named entity recognition, relation extraction, conversational AI, and knowledge graphs. I’ve trained large language and multimodal models with up to seventy billion parameters using QLoRA, distributed training, and pipeline parallelism, including hands-on experience across NoSQL, Graph, and Vector databases.

Experience

Work history, roles, and key accomplishments

Current

AI Research Scientist

Current

Stealth Startup

Oct 2025 - Present (9 months)

Applied research on agentic AI for workspace management in 3D VR, including publishing patents on the work. Deployed a pipeline using MLLMs, custom agentic harness/MCP servers, and LLM evals using cognitive load metrics from eye tracking.

PyTorch Tool Use LLMOps Nvidia DeepStream Varjo XR 4

Doctoral Researcher/Data Scientist

University of Maryland, College Park

Jan 2019 - Aug 2025 (6 years 7 months)

Researched how non-experts in ML can train, test, and deploy image classifiers effectively, resulting in 7 publications, 2 awards, and 3 fellowships. Built and evaluated multiple on-device and web-based tools, including iOS apps and a full-stack dataset repository website.

PyTorch Swift React Node TypeScript MongoDB ResNet Git GraphQL NoSQL iOS ARCore

Research Data Scientist

Fujitsu Research

Jan 2025 - Apr 2025 (3 months)

Researched human-centered multimodal agents for video content moderation, improving individual system F1 from 0.5 to 0.8 and cutting latency by half for long videos. Built and benchmarked multimodal pipelines including MLLMs, detectors/segmentation models, and RAG/knowledge-graph generation components.

PyTorch Tool Use LLMOps Supabase BERT Aim MLFlow LangGraph AWS Sagemaker DVC VLLM Knowledge Graphs

Software Engineer

SplashLearn

Sep 2017 - Jul 2018 (10 months)

Led a team of designers, developers, and artists to launch educational games in the US, resulting in 40+ games. Designed and developed K-5 math teaching games, managed sprint planning/cross-disciplinary collaboration, and created reusable gameplay frameworks and tools.

PyTorch JavaScript Pandas Jira GitHub Refactoring SOLID

Research Data Scientist

IBM Research

Jul 2015 - Aug 2017 (2 years 1 month)

Worked on AI-driven education projects using IBM Watson and its GenAI models, resulting in 3 publications and 5 patents. Trained autoregressive LLMs and fine-tuned BERT models, including usability experiments with eye tracking and AR, achieving up to 0.83 F1 on educational use cases.

PyTorch BERT Pandas NumPy MLOps GitHub Tinkerpop ResNet