Research
Research & Publications
My work spans AI/ML systems, network security, and human-AI collaboration. Open to research opportunities and collaborations.
My work spans AI/ML systems, network security, and human-AI collaboration. Open to research opportunities and collaborations.
Research Assistant under Dr. Laura Globig
Contributing to interdisciplinary research that takes a global perspective, integrating fundamental scientific inquiry with practical application to deliver robust, versatile, and globally scalable insights. The center brings together scientists worldwide to understand the social dynamics of intergroup conflict and cooperation, and their interconnectedness with topics like AI technology, misinformation, climate change, social media, and well-being.
My work involves generating insights using R Studio, analyzing existing literature on AI across developed and developing nations, investigating equity and access disparities, and supporting LLM-based data testing and scaling efforts.
Center for Conflict & CooperationClick any entry to expand
IEEE International Conference PuneCon 2022 · Dec 16, 2022
Between the late-night debugging sessions and the "just one more paper" rabbit holes, it's coming. (I don't have a social life, but I do have citations.)
Click any entry to expand
NE Agents Day 2026 Workshop · Submission #69 · CC BY 4.0
Author: Adrian Dsouza | Submitted: 06 Apr 2026 (modified: 23 Apr 2026)
Keywords: AI agents, product management automation, human-in-the-loop systems, RAG, prompt refinement
TL;DR: AgentFlow is an AI-human orchestration prototype for product management automation. The system combines retrieval-augmented generation, iterative meta-prompt refinement, and dynamic routing between language-model agents and human reviewers.
Abstract: While large language models demonstrate strong performance in isolated tasks, their deployment in real-world workflows remains limited by challenges in state management, tool use, and coordination across multi-step processes. AgentFlow, an AI-human orchestration system designed for product management workflows, addresses this. Given a project brief and a set of retrieved artifacts, the system produces structured outputs including feature breakdowns, user stories, acceptance criteria, and sprint-ready tasks. A central component is iterative prompt refinement, where outputs are progressively improved through feedback from prior generations, retrieval quality, and human edits. Product workflows provide a meaningful testbed for agent research, exposing core challenges in state representation, orchestration policies, and evaluation beyond single-turn accuracy.
NYU Meta-Research Symposium · New York · Poster Presentation
From prototype to poster: AgentFlow, an agentic AI + RAG framework that decomposes project briefs into atomic tasks, injects the right context, and routes work between AI agents and human experts with clear, auditable hand-offs. In meta-research, this translates into more reliable citation screening, cleaner reporting workflows, and stronger reproducibility. Hosted on AWS S3 and EC2.
Inspired by keynote speakers Dr. Ruopeng An, Dr. Elizabeth Stevens, and Dr. Rémi Thériault. Special thanks to Dr. Stevens, Dr. Thériault, Alzahraa (Zahra) K. Ahmed, and Katerina Andreadis for encouragement and feedback.
AWS 10,000 AIdeas Semi-Finalist. A deep dive into AgentFlow, an AI-human orchestration platform that rethinks project management from the ground up. Covers the architecture (RAG, FAISS, Meta-Llama-3, FastAPI, WebSockets), iterative prompt refinement, human-in-the-loop routing, and lessons learned building agentic systems for real-world workflows.
Published in Towards Artificial Intelligence. Explores a multi-agent system that goes beyond lowest-bid hiring by building a deterministic, explainable framework for evaluating freelancers. Covers agent design, scoring rubrics, explainability layers, and how AI can make hiring decisions that are transparent, auditable, and fair.
Builds a smart photo labelling and search system on AWS — combining Rekognition for automatic image tagging, S3 for long-term storage, and custom metadata to create a searchable, intelligent photo backup. Inspired by the smart labelling features in phone gallery apps, extended to a scalable cloud-native architecture.
A walkthrough on architecting a fully automated ML workflow: data ingestion with SageMaker Processing Jobs & Feature Store, hyperparameter sweeps with Automatic Model Tuning, CI/CD with SageMaker Pipelines + CodePipeline, and monitoring with Model Monitor for data drift, concept drift, and bias detection.
Covers Elastic Beanstalk's end-to-end architecture, deploying production-ready applications using EC2, S3, Load Balancer, and SQS, with a real-world photo-to-metadata pipeline. Includes RESTful routing and cost optimization strategies for launching full-stack solutions in under 200ms.
Designing systems where AI agents and human experts co-create, with iterative prompt refinement and task orchestration.
Building on my IDS research, exploring adversarial robustness, anomaly detection, and trustworthy AI systems.
Investigating AI equity across developed and developing nations, informed by my current RA work at NYU.
Cloud-native AI architectures, RAG pipelines, and production ML systems that bridge research and deployment.
This section will grow as I publish more papers and refine my research direction. Open to research collaborations in AI/ML engineering and human-AI interaction.
ORCID