Raquib Bin Yousuf, PhD

• Designed metadata-aware dual-encoder retrieval methods that incorporate structured disambiguation signals into embedding and ranking objectives.
• Improved retrieval performance by 70% on average over text-only baselines in retrieval-augmented generation settings.
• Developed in collaboration with Vectorize.io, with attention to practical retrieval concerns including schema ambiguity, metadata use, source grounding, ranking quality, and evaluation.

• Built memory-augmented architectures for LLM-based reasoning across multiple documents, using persistent context representations and cross-document evidence linking.
• Combined retrieval, structured memory, and generation to support analytical reasoning over long, distributed information spaces.
• Achieved 25% relative improvement in classification performance and 50% improvement in generation quality for multi-document analytical reasoning tasks.

• Developed a graph-based benchmark to evaluate whether LLMs can maintain stable structured memory as context length increases.
• Studied how models induce, update, and preserve graph-like representations over long contexts, exposing reliability failures that are difficult to detect with standard long-context tests.
• Revealed memory-drift onset at only 1–6% of advertised context capacity, showing that nominal context length can substantially overstate reliable reasoning capacity.

• Built an interactive human–LLM system for exploratory structured data analysis and analytical reasoning over complex datasets.
• Designed workflows for schema understanding, user-guided analysis, grounded generation, and iterative refinement.
• Enabled a human-in-the-loop agentic structured data analysis system with 86% schema-alignment reliability across hundreds of newsroom sessions.
• Developed in collaboration with The Chronicle of Higher Education as a deployed system and research platform for studying AI-assisted data analysis.

• Fine-tuned and evaluated LLMs on AWS EC2 and SageMaker for newsroom applications including subheadline generation, summarization, and question answering.
• Contributed to the early development of the Ask the Post chatbot, advising on RAG design, grounding, and evaluation in collaboration with newsroom stakeholders.
• Presented LLM evaluation, fine-tuning, and model development work to engineering leadership and newsroom audiences.
• Focused on editorial quality, reliability, answer grounding, and model behavior in newsroom use cases.

• Led development of ML systems for product provenance verification, combining probabilistic modeling, data valuation, and spatiotemporal reasoning for compliance workflows.
• Supported assessment of 59 wood products, identification of 260+ tons of allegedly illegal timber, and 9+ enforcement investigations.
• Developed in collaboration with World Forest ID on real-world traceability problems involving noisy data, uncertain labels, and regulatory decision-support needs.
• Contributed to research on optimizing product provenance verification using data valuation methods.

• Built large-scale forecasting pipelines for migration patterns and land border encounters using applied ML, statistical modeling, and spatiotemporal data.
• Developed real-time and policy-facing predictive workflows designed for decision support under uncertainty.
• Worked on projects funded by government and research partners, emphasizing scalable pipelines, evaluation, and decision-support relevance.

• Designed a full-text scientific information extraction system using domain-adapted transformer models and task-specific representation learning objectives.
• Achieved 11% improvement over prior baselines and 26% higher accuracy in salient task and method extraction from scholarly documents.
• Worked with collaborators at CSET, Georgetown University, to extract structured signals from scholarly documents for science-of-science analysis.

• Introduced permutation-aware tabular data generation methods to reduce invalid synthetic data generation by LLMs.
• Reduced synthetic table rule violations by 70% through structured generation constraints and evaluation.
• Contributed to Fisher information-guided regularization for language model fine-tuning, improving generalization in low-data regimes across 9/10 GLUE tasks with no added computational overhead.

• Build data-grounding layers for heterogeneous traffic and mobility data by standardizing schemas, metadata, and data-access patterns.
• Develop LLM-based traffic analysis tools that answer natural-language questions through grounded querying, analysis, and summaries.
• Create benchmarks to evaluate and improve the reliability of AI tools over complex traffic datasets and external context.

• Fine-tuned and evaluated LLMs on AWS EC2 and SageMaker to explore newsroom applications including subheadline generation, summarization, and question answering.
• Contributed to early development of the Ask the Post chatbot, advising on RAG design and evaluation in collaboration with newsroom stakeholders.
• Presented LLM evaluation, fine-tuning, and model development for newsroom applications to engineering leadership and newsroom audiences.

Built LLM, applied ML, NLP, retrieval, and forecasting systems for data-aware reasoning and decision support across funded research projects. Selected work:

• Developed memory-augmented architectures for LLM-based multi-document reasoning with persistent context representation and cross-document evidence linking.
• Proposed metadata-aware retrieval methods for RAG, incorporating structured disambiguation signals into embedding and ranking objectives in collaboration with Vectorize.io.
• Built human-in-the-loop agentic data analysis workflows for real-world newsroom use cases in collaboration with The Chronicle of Higher Education.
• Introduced permutation-aware tabular generation and Fisher information-guided regularization for LLM fine-tuning to improve data efficiency and generalization.
• Designed full-text scientific information extraction systems using domain-adapted transformer models and task-specific representation learning.
• Led development of ML systems for product provenance verification and large-scale migration forecasting, deploying regulatory and real-time pipelines used in policy and compliance settings, including provenance work in collaboration with World Forest ID.
• Contributed to funded research proposals for projects supported by DARPA, NSF, and external partners, including technical section writing.

• Led advanced programming and digital logic design courses, managed labs, exams, student supervision, and administrative responsibilities.
• Served on academic and administrative committees and contributed to departmental teaching and curriculum activities.