Setareh Rafatirad

My current research focuses on AI Agents, Applied Machine Learning, Foundational Models, Personalization, Reasoning, Low-Resource NLP, and AI in Healthcare. I am particularly interested in building robust, interpretable, and human-aligned AI systems that address real-world challenges. Some of the ongoing and completed projects are detailed below.

This project focuses on designing a deep learning-based traffic prediction and management system for California highways. The system leverages historical traffic volumes, sensor feeds, and external contextual factors such as holidays, weather, and events. We employ spatiotemporal deep learning architectures including LSTM networks and graph-based models to capture both temporal trends and spatial correlations across highway segments.

The ultimate objective is to support dynamic traffic rerouting, signal coordination, and emergency response planning. In addition, we explore explainable ML techniques to ensure interpretable outputs that assist transportation planners and Caltrans engineers in real-time operational decision-making.

This project introduces a health-tracking wearable integrated with conversational AI that monitors vital signals such as heart rate, temperature, and sleep quality. The device is coupled with a personalized AI assistant that can respond to health queries in natural language, such as - I have had a headache for two days, what should I do?

We employ transformer-based LLMs to generate responses that account for tracked biometrics and medical symptom models. This fusion of physiological monitoring with AI interaction creates a powerful tool for early detection, chronic condition management, and digital health equity.

This project introduces a novel perturbation-based benchmark for evaluating the robustness of Large Language Models (LLMs) when classifying bias in public policy statements. Starting with a curated dataset from federal and state-level sources, we label excerpts for racial, gender, or economic bias categories.

We apply a wide range of perturbations - syntactic reordering, gibberish injection, adversarial rewrites and measure classification degradation across 7 state-of-the-art LLMs. The resulting framework reveals model brittleness under minimal textual changes, informing both model developers and policy analysts.

We explore the design and implementation of autonomous AI agents that can break down complex goals into executable tasks, use external tools and APIs, and collaborate with humans through natural conversation. Our agent architecture integrates LLMs with memory modules, planning routines, and action execution.

Use-cases include literature review automation, project planning assistance, and research workflow management. The project contributes to the future of reliable human-AI partnerships in both academic and industrial settings.

This research addresses the challenge of building LLMs that can reason over text, images, and structured data simultaneously. We develop synthetic and real-world benchmarks that test capabilities like visual entailment, multimodal QA, and temporal inference.

By enhancing alignment across modalities through attention mechanisms and joint embedding spaces, we aim to support applications such as medical report generation, instructional video analysis, and AI tutors.

As LLMs are increasingly deployed in real-world applications, traditional metrics like accuracy are insufficient to capture their reliability. This project develops a holistic evaluation suite including robustness to perturbations, fairness across demographic groups, factual consistency, calibration, and adversarial stress-resistance.

Our benchmark enables developers and researchers to perform nuanced assessments of LLMs before production deployment, especially in education, healthcare, legal reasoning, and public policy domains.

Personalization is crucial for user-facing systems such as learning platforms, digital assistants, and recommender engines. In this project, we develop learning pipelines that adapt to individual users through continual feedback, preference modeling, and on-device learning.

We explore federated learning for privacy-preserving personalization and memory-augmented neural networks for long-term adaptation, ensuring scalability and responsiveness without compromising data confidentiality.

This project investigates the use of foundation models, pre-trained on large corpora, for accelerating scientific discovery. We fine-tune and adapt these models for domain-specific tasks in areas such as climate forecasting, materials synthesis, and policy analysis.

Our approach emphasizes data efficiency, knowledge-grounded generation, and interpretability. Outputs include automated hypothesis generation, structured summarization of scientific papers, and machine-generated experimental design suggestions.

This research aims to apply machine learning techniques to real-world societal challenges such as urban planning, agriculture, public health, and disaster relief. We utilize data from satellite imagery, government records, IoT devices, and social signals to build predictive tools.

Emphasis is placed on scalability, interpretability, and ethical design. The resulting systems are deployed in collaboration with public-sector agencies and NGOs to assist in early warnings, resource allocation, and infrastructure resilience.

NLP technologies have predominantly focused on high-resource languages, leaving billions of speakers of low-resource languages underserved. In this project, we build models for translation, sentiment analysis, and information retrieval for underrepresented languages.

Through transfer learning, multilingual embeddings, and community-based dataset creation, we aim to democratize AI and promote linguistic inclusion in digital systems. Our models are designed to be lightweight, interpretable, and easily deployable.

Data Science has recently experienced a significant surge in producing undergraduate degree/certificate and course enrollments. This growth has resulted in straining program resources at many institutions and causing concern about how to most effectively respond to the rapidly growing demand and provide equal opportunities for all students to pursue their degree. The primary goal of this work is to provide an intelligent solution to automatically identify biases related to race and gender in the content of Data Science related courses such as Machine Learning (ML), Natural Language Processing (NLP), Data Mining, Data Analytics, and Knowledge Discovery just to name a few. While machine learning allows to build powerful predicting tools, it hasn’t been used sufficiently to detect biases rooted in gender and race in educational material.

Hence, using effective machine learning techniques we aim to enhance the success rate of undergraduate students in Data Science education and programs while promoting engagement through identifying and mitigating potential gender and racial biases. To this end, we propose a novel bias detection approach which uses a combination of gender identification and sentiment classification for characterization of gender and race in educational contents. Our proposed framework can help educators and course developers to mitigate race and gender biases in their course materials and create an equitable learning experience for students of minority and under-represented groups.

Stealthy malware is a malware created by embedding the malware in a benign application through advanced obfuscation strategies to thwart the detection. To perform efficient malware detection for traditional and stealthy malware alike, we propose a two-pronged approach. Firstly, we extract the microarchitectural traces obtained while executing the application, which are fed to the traditional ML classifiers to detect malware spawned as separate thread. In parallel, for an efficient stealthy malware detection, we introduce an automated localized feature extraction technique that will be further processed using the recurrent neural networks (RNNs) for classification. To perform this, we translate the application binaries into images and further convert it into sequences and extract local features for stealthy malware detection.With the proposed two-pronged approach, an accuracy of 94% and nearly 90% is achieved in detecting normal and stealthy malware created through code relocation obfuscation technique. Furthermore, the proposed approach achieves up to 11% higher -

detection accuracy compared to the CNN-based sequence classification and hidden Markov model (HMM) based approaches in detecting stealthy malware.The results of this work has been published in DATE 2019, ICTAI 2019, ICMLA 2019, CASES-ESWEEK 2019 conferences.

Malware detection at the hardware level is an effective solution to increasing security threats. Hardware based detectors rely on Machine-Learning (ML)classifiers to detect malware-like execution pattern based on Hardware Performance Counters(HPC) information at run-time. The effectiveness of these methods relies on the information provided by expensive-to-implement limited number of HPCs. We analyze various robust ML methods to classify benign and malware applications and effectively improve the classification accuracy by selecting important HPCs. We fully implemented these classier at OS Kernel to understand various software overheads. The software implementation of these classifiers found to be relatively slow.

We propose a hardware-accelerated implementation of these algorithms where the studied classifiers are synthesized to compare various design parameters including logic area, power, and latency. The results show that while ML classier such as MLP are achieving close to 90% accuracy, after taking into consideration their implementation overheads, they perform worst in terms of PDP, accuracy/area and latency compared to simpler but less accurate rule based and tree based classifiers. Results show OneR to be the most cost-effective classier, achieving highest accuracy per logic area, while relying on only branch-instruction HPC information. This work has led to a publication in DAC 2017 conference and a journal in ACM TODAES.

Scale-out infrastructure such as Cloud is built upon a large network of multi-core processors. Performance, power consumption, and capital cost of such infrastructure depend on the overall system configuration including number of processing cores, core frequency, memory hierarchy and capacity, number of memory channels, and memory data rate. Among these parameters, memory subsystem is known to be one of the performance bottlenecks, contributing significantly to the overall capital and operational cost of the server. Also, given the rise of Big Data and analytics applications, this could potentially pose an even bigger challenge to the performance of cloud applications and cost of cloud infrastructure. Hence it is important to understand the role of memory subsystem in cloud infrastructure. Currently there is no well-defined methodology for selecting a memory configuration that reduces execution time and power consumption by considering the capital and operational cost of cloud.

In this work, through a comprehensive real-system empirical analysis of performance, we address these challenges by first characterizing diverse types of scale-out applications across a wide range of memory configuration parameters. The characterization helps to accurately capture applications’ behavior and derive a model to predict their performance. We propose MeNa, which is a methodology to maximize the performance/cost ratio of scale-out applications running in cloud environment. MeNa navigates memory and processor parameters to find the system configuration for a given application and a given budget, to maximum performance. Compared to brute force method, MeNa achieves more than 90% accuracy for identifying the right configuration parameters to maximize performance/cost ratio. This work has led to a conference publication in IEEE IISWC 2017.

The advent of technology and abundance of sensors on handheld devices has led to the proliferation of open source datasets with sensory metadata (like time, location, so forth). Cloud-based off-the-shelf computer vision APIs like Project Oxford and Clarifai are abundantly becoming available to report high-level metadata concepts from the images pixel values. The availability of semantic web technology as defacto standard to create structured metadata and ontologies for a domain of interest, linked-data sources (like Google knowledge graph and Wikipedia) -providing background knowledge, big data analytics and deep learning algorithms,

and various web services and public database APIs, combined, provides semantic richness which should be leveraged to infer high-level semantics and bridge the semantic gap. The goal of this project is integrating the metadata from personal photo streams with information provided by various contextual data sources, APIs, tools, and techniques to model spatio-temporal entities,called events, and construct event-based domain ontologies, which later leads to detecting daily activities and events, and discovering correlations among them. Such outcome can be used in variety of applications including multimedia information retrieval, information organization, and healthcare.

Big data is transforming several domains, including real-estate industry. Real-estate companies like Zillow, Redfin, and Realtor offer applications utilized by expert/ordinary users. However, these applications lack the necessary components for providing a personalized analytical framework to meet the individual need to explore and choose the best investment options. The goal of this project is to address this problem. In our research group, a team of students are currently developing the code and software to collect, analyze, and visualize housing data and environmental factors for the U.S. regions. We are exploring a variety of Machine Learning and Deep Learning techniques for housing price prediction, and improve the accuracy of the results. In this project, we aim to improve the accuracy of our prediction algorithm.

We have tested several traditional ML algorithms on our data. We will examine Ensemble Learning (such as Random Forest) and Deep Learning techniques (such as CNN) as the size of data grows. In addition, we have collected historic data for all US states which we are using to develop time-series rent prediction model. However, with growing size of data, we are facing several challenges including high frequency of missing data. We will investigate an array of imputation techniques to avoid from a biased training set, and therefore, end up with higher model accuracy. In addition, we are investigating many external parameters that could influence the prediction models related to urban development. Another goal of this ongoing project is creating a mobile app software for both Android and iOS devices, so that system users can visually and empirically interact with the results of our research findings.