H.A.R.M.O.N.I. Lab

Human-Aligned, Resilient, Multimodal, Open-ended, Novelty-informed Intelligence

Director: KMA Solaiman

Designing AI That Endures

At the H.A.R.M.O.N.I. Lab, we design AI systems that think beyond narrow objectives — systems that adapt, align, and reason in complex, real-world environments. From smart grids to policy documents, from missing persons to triage prediction, our research spans multiple modalities and domains, bridging real-world applications with foundational AI challenges.

Our work is organized into three core clusters, each linked to active 📌 projects and real-world applications:

Applications We Explore

Lab Projects

Novelty-Aware Smart Electric Grids

Designing a predictive divergence–based novelty detection and adaptive planning framework for smart electric grids. This system guides real-time control actions in uncertain operational scenarios and performs triage.
Current Researchers: KMA Solaiman (PI), Sumya Hamid Oishi
Related Publications:

Multimodal Information Retrieval and Alignment

This project line explores how intelligent systems retrieve, align, and reason over multimodal information — including images, documents, and video — using graph-based, hierarchical, and layout-aware techniques. The focus spans structural matching, human-in-the-loop retrieval, and open-world adaptation across diverse content modalities.
Current Researchers: KMA Solaima (PI)
Related Publications:

Structure-Aware Novelty in Open-World AI

This project develops foundational methods to model, quantify, and adapt to novelty in intelligent systems. We study how classifiers behave under distributional shift, how domain complexity affects generalization, and how structural representations can support uncertainty-aware learning. Techniques span embedding-based signal fusion, selective abstention, and severity scoring for downstream response.
Current Researchers: KMA Solaiman, Shafkat Islam
Related Publications:

Motherboard Defect Detection

Student-led project exploring visual anomaly detection in motherboard hardware using deep learning. The system applies YOLOv7 to classify and localize defect regions from inspection images captured in lab conditions.
Current Researchers: KMA Solaiman, Brandon Hill
Related Publications:

Unsupervised Clustering and Structure Discovery

Exploring clustering algorithms that reveal latent structure in unlabeled, high-dimensional data. This work emphasizes robustness to irregular clusters and scale variance.
Current Researchers: KMA Solaiman
Related Publications:

Human-Centered Reasoning & Recommendation Systems

These projects center around human-aligned AI systems designed to reason with real-world constraints and preferences. From public safety analytics to personalized consumptions (i.e. plants, content) matching, the focus is on making decision-support tools accessible, interpretable, and adaptable to user context.
Current Researchers: KMA Solaiman
Related Publications:

Adaptive Prediction in Triage & Financial Markets

This project explores how multimodal learning can improve decision-making in high-stakes domains like medical triage and stock market prediction. We model temporal and semantic signals — including vitals, news, and sentiment — to forecast outcomes under uncertainty. The research emphasizes cross-domain learning and robust performance on evolving, real-world data streams.
Current Researchers: KMA Solaiman (PI), Adam Sayeed, Joshua Sebastian, Karma Tobden
Related Publications:

BiasLab: Explainable Political Bias Detection

We introduce a human-annotated dataset for studying perceived political bias in news headlines and rationales. The project examines framing effects, subjective disagreement, and annotation typologies. We compare human judgments with LLM outputs to expose bias perception gaps and develop explainable modeling pipelines.
Current Researchers: KMA Solaiman
Related Publications:

Course-Based Research Topics (CMSC 678 / CMSC 471 / CMSC 478)

Students in my courses occasionally pursue independent or group research projects that align with the H.A.R.M.O.N.I. Lab’s core research clusters.
These topics are examples of open project directions that students may select or adapt for future semesters.
Each will be scoped appropriately for course credit and designed around public data and independent code.

🔹 Causal Summarization of Safety Narratives

Cluster: Multimodal Interpretability & Reasoning (H.A.R.M.O.N.I. Cluster 3)
Explore how large language models (LLMs) can summarize and extract causal or contributing factors from publicly available safety reports or incident narratives.
Focus: prompting design, summarization quality, and interpretable causal reasoning.
(Uses public datasets only; conducted under H.A.R.M.O.N.I. Lab supervision for educational and exploratory research.)

🔹 Novelty Detection in Smart-Infrastructure Sensor Data

Cluster: Resilient & Novelty-Aware AI Systems (H.A.R.M.O.N.I. Cluster 1)
Investigate methods for identifying unusual events or rare patterns in open traffic-sensor datasets using independently developed code.
Focus: novelty detection, temporal patterning, and visualization of rare behaviors.
(Independent codebase and open data only.)

🔹 Survey + Evaluation of Sensor–Text Embedding Spaces

Cluster: Multimodal Representation Alignment (H.A.R.M.O.N.I. Cluster 2)
Examine how numerical sensor data (e.g., EEG or motion signals) and natural-language descriptions can be represented and compared in a shared latent space.
Focus: cross-modal embedding geometry, metric robustness, and visualization of alignment patterns.
(Uses public time-series datasets; conducted under H.A.R.M.O.N.I. Lab supervision for educational and exploratory research.)

✳️ Attribution & Ethics

All course projects must use publicly available datasets and independently written code.
Outstanding work may later be extended for academic publication under H.A.R.M.O.N.I. Lab supervision with appropriate student credit or acknowledgement. Any resulting models, embeddings, or visualizations may not be reused externally without permission from instructor.

Join the Lab

We welcome students who are curious, motivated, and eager to build practical AI systems that tackle real-world complexity. Your time is valuable, and we strive to ensure that your contributions are recognized — through course credit, funding, or formal research roles.

Who We’re Looking For

For UMBC Master’s and Undergraduate Students

While I currently do not have dedicated funding for master’s or undergraduate students, I am open to supervising independent projects and collaborative research for credit.

I strongly prefer to get to know students through coursework before working together. If you’re interested in joining the lab, please consider enrolling in one of my classes.

While I would like to work with many exceptional students, I may not always be able to accommodate everyone. However there are several useful resources and programs at UMBC:

Courtesy: List based off of Dr. Tejas Gokhale’s FAQ.

Collaborations

We collaborate across campus and beyond:

Contact

KMA Solaiman
Assistant Teaching Professor, CSEE, UMBC
Director, H.A.R.M.O.N.I. Lab
📧 ksolaima@umbc.edu
🌐 Lab Website

If you’re driven by curiosity and care about the real-world impact of AI, we’d love to hear from you.