CV Download PDF

Masters in Computer Science

• Graduated with Computer Science Masters at Johns Hopkins University.

BS Computer Science, Applied Mathematics & Statistics

• GPA: 3.90/4.0
• Course Assistant for NLP: Self-supervised Models, Natural Language Processing, Introduction to Algorithms, Automata and Computation Theory, Discrete Math.
• Math Tournament Problem Writer: ASDAN, JHMT (Johns Hopkins Math Tournament).

Afterglow: Forward Span Propagation for Query-Aware KV Cache Eviction

• Improves published query-aware KV cache eviction by propagating high attention score with decay to later tokens.
• Currently scores first in kvpress leaderboard for query-aware single pass eviction method.

VFold: Symmetry-Aware Cross-Layer Value Cache Compression

• Inter-layer cache merging algorithm that exploits cross-layer redundancy.

Sparse Attention with Entropy-Triggered KV Repair

• Entropy triggered sparse attention cache rebuild for accurate responses.

Recovering Collapsed Language Models in Weight Space Without External Data

• Experimented multi-model feedback loop and confirmed that model collapse appears in this scenario.
• Found that collapse is locally reversible via task vector negation.

What Survives Eviction? Information Recall Locality in Transformer KV Caches

• Explored cache locality in information retrieval.

BiomedBench Suite: Benchmarks for Evaluating LLM Performance on Biomedical Reasoning Tasks

BiomedSQL: Text-to-SQL for Scientific Reasoning on Biomedical Knowledge Bases

• Dataset for evaluating scientific reasoning in text-to-SQL generation over a real-world biomedical knowledge base.

CARDBiomedBench: A Benchmark for Evaluating Large Language Model Performance in Biomedical Research

• Large Scale Question and Answer Benchmark for evaluating LLMs in biomedical science.

Challenging the Evaluator: LLM Sycophancy Under User Rebuttal

• Investigated qualities that trigger LLM sycophancy under user rebuttal.

Research Assistant

• Research Assistant under Professor Daniel Khashabi

Part Time

• Engineered a medical RAG(Retrieval Augmented Generation) pipeline. Implementing various methods for SQL retrieval including chain-of-thought reasoning and self-consistency.
• Helped create datasets and evaluation metrics for medical questions.

Research Intern

• Worked on natural language to SQL query pipeline that scales sublinearly to the number of columns
• Achieved 5% increase in accuracy (80% to 85%) compared to vanilla schema-as-context methods.

Research Assistant

• Created chatbots which cite databases for a web based app.
• Developed and tested various retrieval methods including dense, sparse, hybrid retrieval methods, chunking strategies, LLM fine tuning, for robust search using Huggingface, sbert, Apache Solr, nltk, and more.

Research Assistant

• Created neural network-derived brain perfusion map processing pipeline for model-free CTP images.
• Achieved up to 30% better MSE loss compared to a recent paper on the same dataset.

Information Systems Maintenance/Operator, Squad Leader

• Maintained and operated secure military information systems to ensure reliable communication of sensitive data.
• Appointed Squad Leader in June of 2021.

Platform Analyst (Intern)

• Analyzed game data logs using SQL, Hadoop, and Python.
• Built ML model (Random Forest, LSTM, XGBoost) to classify player experience levels, improving accuracy by 15%+ over a Markov Chain baseline.

Logit Lens Llama Advanced

• Created an interactive version of Logit Lens, where users can intervene on residual, attention and FFN stream, on a per-layer, per-head basis.

Embedding-Level Translation for Seamless Communication Across LLM Agents

• Developed an embedding-level translator to enable seamless communication between LLM agents, achieving <0.01 MSE and >0.9 cosine similarity, demonstrating high-fidelity translation across agent representations.

Linear Embedding Space Alignment for Semantic Change Detection

• Explored limitations of the Procrustes alignment method for detecting domain shifts in semantic spaces over time.
• Proposed and evaluated alternative linear alignment techniques using unsupervised anchor discovery, tested on a machine translation benchmark.
• Found that while methods varied in approach, no significant performance difference was observed across alignment strategies.

arXiv Reference Finder (Hophacks - 2nd Place Winner)

• Built a tool to extract and highlight relevant sections from referenced papers on arXiv, given a source paper.
• Utilized BM25 and cross-encoders for semantic similarity; integrated arXiv API and applied regex-based PDF chunking and formatting.

Custom Transformers

• Implemented transformer layers with non-standard architectures.
• Achieved lower classification loss than HuggingFace BERT, despite using fewer parameters.

Misc Game Tools (Toys)

• Designed an Excel-based Markov chain model to calculate expected costs of in-game equipment upgrades.
• Developed a course planner to track and store all Johns Hopkins University course data locally.
• Created a Pokémon battle damage calculator and derived a custom method for type ranking analysis.