cv

Ph.D. in Computer Science

Howard University

B.S. in Computer Science

Howard University

Machine Learning Researcher

Artificial Intelligence for Positive Change Lab @ Howard University

Advanced research on large language model adaptation, alignment, and instruction tuning for clinical NLP, achieving state-of-the-art ADE detection and leading cross-lab collaborations on multimodal and agentic AI systems.

• Advanced research on LLM adaptation, supervised fine-tuning, and instruction tuning, achieving state-of-the-art ADE detection across biomedical corpora.
• Improved LLM alignment and robustness using RLHF and DPO, reducing factual inconsistency and overprediction by 30% in clinical NLP.
• Led a team of five researchers in developing multimodal fusion models and agentic reasoning systems integrating speech, text, and knowledge-grounded inference, with publications in ACL, IEEE, AAAI, and PSB.
• Directed the National Geospatial-Intelligence Agency (NGA) team on multimodal and geospatial AI, delivering ML solutions for GIS and signal-based data analysis.

Machine Learning Engineer Intern

Reddit

Designed and implemented a novel time-aware Transformer architecture introducing long-horizon engagement forecasting, achieving 85% prediction accuracy and Reddit’s first production-scale engagement forecast model.

• Designed and implemented a time-aware Transformer architecture for long-horizon engagement forecasting across billions of interactions, achieving 85% prediction accuracy.
• Built a distributed training and inference pipeline (PyTorch Lightning + Ray) reducing training latency by 40% and enabling a 600× faster inference pipeline for near real-time forecasts.
• Conducted LLM interpretability analyses via attention and embedding probing, deploying learned embeddings as user representations to enhance personalization and recommendation quality.

Machine Learning Engineer Intern

Reddit

Developed an LLM-driven onboarding recommender system leveraging retrieval-augmented generation and structured prompting, reducing cold-start friction by 25% and improving recommendation variety through multi-agent orchestration.

• Developed an LLM-driven onboarding recommender system using RAG and structured prompting, reducing cold-start friction by 25% in internal A/B testing.
• Designed a prompt orchestration layer (Go + gRPC + GraphQL) for multi-agent content workflows, enabling scalable real-time personalization.
• Built LLM-as-a-Judge evaluation pipelines combining human preference scoring and reward modeling, improving recommendation accuracy by 14%.

Software Engineer Intern

Reddit

Built developer SDKs and APIs for Reddit’s Developer Platform, improving integration workflows and adoption across internal and external teams.

• Developed React/TypeScript SDKs and REST APIs for Reddit’s Developer Platform, enabling third-party integrations and adoption by over 300 developers.
• Created frontend tools improving usability, documentation, and integration workflows across partner engineering teams.
• Optimized API performance and scalability in collaboration with backend teams, improving response latency and platform reliability.

Undergraduate Machine Learning Researcher

Human Centered Artificial Intelligence Institute @ Howard University

Led multilingual NLP, computer vision, and acoustic analysis projects achieving state-of-the-art performance in code-switched and low-resource tasks, resulting in publications at ACL, ICLR, and PSB.

• Led multilingual NLP, computer vision, and acoustic analysis projects achieving state-of-the-art performance in code-switched sentiment analysis, language identification, and dementia MMSE prediction.
• Designed language-specific transformer fusion systems that improved multilingual classification accuracy by 21%, leading to publications at ACL (SemEval), ICLR, and PSB.
• Organized and taught an undergraduate NLP Bootcamp covering model interpretability, evaluation design, and reproducible ML workflows.

Machine Learning Research Intern

Excella

Built scalable ML experimentation frameworks and synthetic data generation pipelines leveraging GPT and BERT embeddings, enhancing data efficiency and model robustness.

• Built and deployed cloud-native ML experimentation frameworks (AWS/GCP) reducing development time and supporting distributed experimentation.
• Designed synthetic data augmentation pipelines leveraging GPT and BERT embeddings, improving classification robustness by 18%.
• Automated dataset normalization and benchmarking workflows to enhance reproducibility and data quality across ML deployments.

Google PhD Fellowship Recipient

Google

The Google PhD Fellowship recognizes exceptional graduate students conducting innovative and impactful research in computer science and related fields, with the NLP Fellowship specifically awarded for outstanding contributions to Natural Language Processing.

NSF Fellowship Honorable Mention

National Science Foundation

The NSF Fellowship Honorable Mention recognizes promising graduate students who are pursuing research-based master's and doctoral degrees in NSF-supported science, technology, engineering, and mathematics disciplines.

AIM-AHEAD Research Fellowship Recipient

AIM-AHEAD

The AIM-AHEAD Research Fellowship supports researchers leveraging AI/ML to advance health equity, fostering innovation, diversity, and collaboration in addressing healthcare disparities.

Leveraging Large Language Models for Adverse Drug Event Detection: A Comparative Study of Token and Span-Based Named Entity Recognition

Pacific Symposium on Biocomputing (PSB)

Investigates fine-tuned large language models for token and span-based NER in Adverse Drug Event detection using the n2c2 Track 2 dataset. Finds token-based models outperform span-based ones, with ensemble methods like majority voting and XGBoost aggregation improving end-to-end relation extraction, enhancing clinical NLP reliability for patient safety.

Entity Only vs. Inline Approaches: Evaluating LLMs for Adverse Drug Event Detection in Clinical Text (Student Abstract)

Proceedings of the AAAI Conference on Artificial Intelligence

This evaluates prompting strategies for ADE detection on n2c2 clinical text; shows an entity-only extraction approach outperforms inline in precision, recall, and token efficiency.

Chicken Disease Image Classification Using Modern CNNs and Vision Transformers

Information Systems for Business Management (ISBM)

Evaluates modern CNNs (EfficientNet, ConvNeXT) and Vision Transformers (DeiT, Swin Transformer) for classifying poultry diseases from fecal images. Highlights class imbalance challenges and shows deep architectures outperform traditional ML baselines, improving early disease detection for agricultural applications in resource-limited environments.

How State Space Machines can help African Speech Language Identification

IEEE AFRICON

Explores advanced deep learning architectures, including the Audio Spectrogram Transformer (AST) and AfriMamba state space model, for identifying African languages from audio. Demonstrates AfriMamba achieves 97.9% accuracy with nearly double the training speed and six times faster inference than AST, highlighting efficiency gains for multilingual speech processing.

Evaluating Llama-3.1 for Adverse Drug Event Entity and Relationship Extraction Across Prompting Techniques

International Conference on Advances in Computing Research

Assesses Llama-3.1 for ADE detection, NER, and relation extraction using zero-shot, few-shot, and tree-of-thought prompts; proposes a consistency + thought-merging variant and compares methods via F1, precision, and recall.

Ad-hoc Ensemble Approach for Detecting Adverse Drug Events in Electronic Health Records

Consortium For Computing Sciences In Colleges Eastern Regional 2024

Uses zero-shot LLMs combined with ensemble strategies to identify and classify ADEs in EHR clinical notes; experiments compare ensemble techniques and model performance.

Benchmarking Current State-of-the-Art Transformer Models on Token Level Language Identification and Language Pair Identification

International Conference on Computational Science and Computational Intelligence (CSCI)

This study benchmarks state-of-the-art transformer models for token-level language identification (LID) and introduces a novel Language Pair Identification (LPI) task. The results highlight the effectiveness of multilingual transformers in analyzing code-switched data, establishing new baselines for language identification in mixed-language corpora.

Ensembling and Modeling Approaches for Enhancing Alzheimer's Disease Scoring and Severity Assessment

International Conference on Computational Science and Computational Intelligence (CSCI)

This work investigates ensemble modeling techniques to improve the scoring and severity assessment of Alzheimer's Disease. Utilizing computational paralinguistics and predictive models, the research enhances reliability in assessing cognitive decline through machine learning approaches.

Evaluating Ensembled Transformers for Multilingual Code-Switched Sentiment Analysis

International Conference on Computational Science and Computational Intelligence (CSCI)

This research explores the effectiveness of ensembled transformer models in multilingual code-switched sentiment analysis. By combining multiple transformer-based architectures, the study evaluates performance improvements and robustness in handling mixed-language text, showcasing advancements in language modeling for code-switching tasks.

Baselining Performance for Multilingual Code Switching Sentiment Classification

Consortium For Computing Sciences In Colleges Eastern Regional 2023

In multilingual communities, significant social media content contains code-switched data; our research fine-tunes models across five datasets with English code-switched with other languages, outperforming current multi-language pair models on binary sentiment classification and showing comparable results on ternary tasks.

Term Frequency Features vs Transformers: A Comparison for Sentiment Classification of African Languages

Consortium For Computing Sciences In Colleges Eastern Regional 2023

Given the underrepresentation of over 2000 African languages in NLP, the recent release of AfriSenti-SemEval Shared Task 12 provides essential sentiment analysis datasets; our Delta TF-IDF approach shows promise in this low-resource setting, outperforming data-heavy transformer models.

Zero-Shot Classification Reveals Potential Positive Sentiment Bias in African Languages Translations

International Conference on Learning Representations 2023

Using the AfriSenti-SemEval dataset for sentiment analysis on 13 African languages, we translate each to English and employ a BART model for zero-shot classification, finding a potential translation bias towards positive sentiments in African languages.

Feature Importance Analysis for Mini Mental Status Score Prediction in Alzheimer Disease

International Conference on Learning Representations 2023

Using the SHapley Additive exPlanations (SHAP) method, this study proposes predictive models to forecast MMSE scores with 54 key features from a leading model, highlighting the Automated Readability Index (ARI) as the most influential feature for capturing language impairments in dementia patients.

Howard University Computer Science at SemEval-2023 Task 12: A 2-Step System Design for Multilingual Sentiment Classification with Language Identification

The 17th International Workshop on Semantic Evaluation (SemEval-2023) at ACL

The release of AfriSenti-SemEval Task 12 introduced 14 datasets for African languages sentiment analysis; our two proposed approaches, Delta TF-IDF and Language-Specific Model Fusion, showed comparable or superior performance to state-of-the-art models like AfriBERTa, AfroXLMR, and AfroLM.

Sentiment Analysis for Multiple African Languages: A Current Benchmark

Social Impact of AI for Africa (SIAIA-23) at AAAI-23

Despite the growth in sentiment analysis research, African languages remained underexplored until the release of AfriSenti-SemEval Shared Task 12. Our benchmarks across 12 languages revealed that while more data improves per-language model performance, models tailored for African languages excel, and a one-size-fits-all model approach isn't ideal, especially for languages with fewer samples.

Acoustic-Linguistic Features for Modeling Neurological Task Score in Alzheimer’s

Pacific Symposium on Biocomputing 2022 (PSB)

With global life expectancy rising due to medical advancements, there's an urgent need for technologies that detect aging-related cognitive diseases, especially Alzheimer's. Using natural language processing and machine learning, we analyzed speech patterns in Alzheimer's patients and, by comparing ten regression models and over 13,000 features, we outperformed existing baselines in predicting Mini-Mental Status Exam scores, emphasizing the importance of handcrafted linguistic features over acoustic ones.

Sentiment Classification of Code-Switched Text using Pre-Trained Multilingual Embeddings and Segmentation

8th International Conference on Natural Language Computing (NATL 2022)

In our increasingly bilingual world, most natural language processing research remains focused on singular languages. We've developed a sentiment analysis algorithm for mixed-language texts that outperforms baselines by over 11% in accuracy, using semantic similarity from pre-trained models and a tailored word set, offering potential for broader multi-language application.