🔬 Research & Technical Projects

Open Recommendation Benchmark for Reproducible Research with Hidden Tests (ORBIT)

Research Assistant — Carnegie Mellon University, Pittsburgh, PA
February 2025 – June 2025

• Designed and developed ORBIT, a unified sequential recommendation benchmark enabling fair and reproducible model comparison with hidden test evaluation.
• Built ClueWeb-Reco, a high-quality, privacy-preserving web recommendation dataset by mapping private U.S. browsing sequences to public webpages via large-scale dense semantic retrieval.
• Standardized data splits and evaluation protocols to improve cross-paper comparability in recommendation research.

Efficient Dense Retrieval with Boundary-Aware Cluster Routing (BACR)

Research Assistant — Carnegie Mellon University, Pittsburgh, PA
September 2023 – May 2024

• Proposed BACR, a cluster-routing model for fine-grained search over cluster-based Approximate Nearest Neighbor (ANN) indices.
• Improved ANN recall by 8.7% over FAISS IVFFlatIP on MSMARCO Web Search.
• Extended DiskANN with multi-node index construction for memory–SSD hybrid large-scale retrieval.

Mixture-of-Experts Neuro-Computational Network on Fusiform Face Area

Undergraduate Research Assistant — The Cottrell Lab, UC San Diego
January 2023 – June 2023

• Trained word, object, and face expert networks to study subordinate-level image classification in the Fusiform Face Area.
• Designed architectures with ResNet-18, custom CNNs, and gating layers; achieved 99.4% expert gating accuracy.
• Optimized OpenCV data pipelines, reducing epoch loading time from 1 hour to 13 minutes.

Question Answering System with Retrieval-Augmented Generation

Carnegie Mellon University
October 2023 – December 2023

• Developed a rule-based polar question detector to guide Mistral-7B in answering both binary and factual questions (team of four).
• Fine-tuned T5-base for question generation on Wikipedia datasets.
• Integrated Sentence-T5 retrieval for retrieval-augmented answer generation.

Transformer-Based User Intent Classification (Amazon MASSIVE Dataset)

University of California San Diego
December 2022

• Fine-tuned BERT-base-uncased across 60 intent classes, achieving 87.32% sentence-level accuracy (team of five).
• Applied UMAP to analyze representation clustering under supervised contrastive, cross-entropy, and contrastive learning objectives.

Image Captioning with CNN–LSTM Encoder–Decoder Architecture

University of California San Diego
November 2022

• Implemented encoder–decoder models using custom CNNs and ResNet-50 with LSTM for MSCOCO caption generation (team of three).
• Achieved 58.04% BLEU-1 and up to 86.53–100% BLEU-4 for best captions.

ZooSeeker — Android Navigation Application

University of California San Diego
March 2022 – June 2022

• Co-developed an Android application for San Diego Zoo navigation using Android Studio.
• Applied Agile methodology for iterative feature development and usability improvement.