Projects

A curated view of research systems, applied builds, and smaller experiments. Start with the strongest case study, then scan the rest by type.

Featured case study

TREC BioGen: Evidence-grounded RAG and LLM evaluation system

Built and evaluated sparse and dense RAG pipelines over 26.8 million PubMed documents, with Qwen-based generation, reranking, and a dashboard for evidence-linked inspection.

Why it matters: Combined BM25, Pyserini, MedCPT, reranking, and feedback-driven analysis into a deployable evaluation workflow and live viewer.
RAGInformation RetrievalLLMsEvaluationBiomedical QA
Corpus
26.8M PubMed documents
Pipeline
Sparse + dense retrieval with reranking
Output
Evidence-linked inspection viewer
Research systems

Retrieval, grounding, and evaluation work

Projects where evidence quality, system behavior, and inspection mattered as much as raw model output.

Multi-hopLLMsRetrievalMedical QA
2025

Agentic and non-agentic multi-hop medical question answering

Built three multi-hop QA systems combining open-source LLMs with Wikipedia and PubMed retrieval, then compared agentic and controlled non-agentic workflows.

Open →
Key result: The Wikipedia-and-PubMed pipeline performed best on the development set and became part of a BioCreative IX publication.
LLMsClinical NLPClassificationEvidence Extraction
2025

Insomnia classification and evidence extraction

Built baseline shared-task systems for identifying insomnia cases and extracting supporting evidence spans from clinical notes.

Open →
Key result: Compared Llama and Qwen variants on both patient-level classification and evidence extraction for a real shared-task benchmark.
NLPSequence LabelingTransformersRL
2024

Biomedical negation detection with Transformers, CRF, and PPO

Fine-tuned BERT-family models for negation cue and scope detection, then expanded the work with CRF, hierarchical, and PPO-based sequence-labeling approaches.

Open →
Key result: Extended the baseline into CRF and PPO-based pipelines to improve minority-label detection and model label dependencies more explicitly.
Applied builds

Product and platform work

Systems shaped by workflows, users, and constraints beyond a paper or benchmark.

Knowledge GraphsRetrievalQuestion AnsweringLangChain
2024

Knowledge-graph and hybrid retrieval for course enrollment QA

Built a question-answering system for course enrollment using RDF and SPARQL alongside LangChain and FAISS for hybrid retrieval.

Open →
Key result: Compared graph-based, vector-based, and hybrid retrieval modes for natural-language enrollment questions.
SecurityAuthenticationIdentityPatent
2021

Government ID-based single sign-on authentication system (patent)

Proposed a single sign-on authentication workflow using government IDs, with a transparent QR-based verification process for secure identity validation.

Open →
Key result: Patent-track concept focused on practical identity verification flow and traceable user validation.
Foundations

Experiments and systems fundamentals

Smaller builds that still show how I think about modeling, performance, and implementation details.

Machine LearningForecastingApplied AI
2021

Flight delay prediction system

Built a predictive ML engine to classify delayed flights and estimate delay duration, with preprocessing and model selection focused on operational reliability.

Open →
Key result: Two-stage modeling flow designed for actionable delay prediction and practical airline-side planning support.
SystemsComputer ArchitectureDeep Learning
2022

Dynamic branch predictor using deep learning

Team project on deep-learning-based dynamic branch prediction to reduce branching cost and improve processor pipeline execution behavior.

Open →
Key result: Built and analyzed a learned branch prediction pipeline that informed our ICICICT 2022 publication.