About
I enjoy working on machine learning problems, but I’m just as interested in the engineering that makes those models useful outside a notebook.
Currently a Software Development Engineer intern at AWS on the Transit Gateway team, working on distributed networking infrastructure. I also hold three research appointments at Johns Hopkins: contributing to HopDoc, a clinical reasoning system trained on 8.3M de-identified patient records at CDHAI; developing early prediction models for Severe Maternal Morbidity using Mamba State Space Models at the Bloomberg School of Public Health; and building multimodal LLM-as-judge pipelines for scientific paper evaluation at the Data Science and AI Institute.
Before Hopkins, I spent two years as a Machine Learning Engineer at IBM Systems Lab, working on anomaly detection, time-series forecasting, ransomware detection, and agentic LLM pipelines. A lot of that work involved taking ideas from experiments and getting them to work reliably on terabyte-scale telemetry. Along the way, I contributed to three IBM Redbooks white papers and became a Gold Author.
I’m pursuing an MS in Applied Mathematics at Johns Hopkins as a Dean’s Master’s Fellow, with an expected graduation in June 2027. I’ve also contributed to 5 peer-reviewed publications across IEEE, EMNLP, Springer LNNS, and IRE. I’m looking for full-time ML Engineer, AI Engineer, and Software Engineer roles starting June 2027.
Current Focus
Building HopDoc, a clinical reasoning system on 8.3M de-identified EHR records for congestive heart failure readmission and mortality prediction at the Center for Digital Health and AI
Early prediction of Severe Maternal Morbidity from longitudinal EHR data using Mamba State Space Models and cost-sensitive learning under extreme class imbalance
Multimodal LLM-as-judge pipelines for academic paper quality screening, benchmarking frontier vs. open-source vision-language models on scientific writing evaluation
Chain-of-thought reasoning evaluation in frontier LLMs, including tampering detection across 120B-235B parameter reasoning models (EMNLP 2026)
Experience
Working on distributed networking infrastructure that routes traffic across customer VPCs, on-premises networks, and Direct Connect gateways. Building backend service features within AWS internal CI/CD and observability tooling.
Contributing to HopDoc, a clinical reasoning system on the Johns Hopkins REACH dataset (8.3M patients, 1.12B structured OMOP events) for CHF outcome prediction. Built and deployed Lark, an AI knowledge platform with RAG-backed document retrieval (PostgreSQL, Azure AI Search, GPT + Claude). Engineered an AACSB faculty qualification analyzer using multi-stage LLM pipelines with consistency validation.
Developing early prediction models for Severe Maternal Morbidity using longitudinal EHR and ecological data. Optimizing Mamba State Space Models for sparse and irregular clinical observations. Applying cost-sensitive learning for rare-event detection under extreme class imbalance.
Fine-tuned a 32B-parameter vision-language model across distributed multi-GPU infrastructure for civil engineering document understanding. Extended context to 256k tokens via progressive RoPE theta scaling. Built a cloud document processing pipeline on AWS EC2 with DynamoDB, chaining DocLayout-YOLO, Google Document AI, and GPT-4o.
Built a multimodal LLM-as-judge pipeline for academic paper quality screening, combining first-page visual analysis, figure inspection, and full-text evaluation across 35 weighted rubric criteria into a normalized quality score. Evaluated on 200 papers (arXiv + PeerRead ICLR 2017). Benchmarking open-source alternatives with Qwen2.5-VL-7B served via vLLM.
Machine Learning (EN.601.475/675) and AI-Enabled Software Engineering (EN.601.423/623). Led recitations on neural networks, RL, and causal inference. Mentored 3 graduate student teams through end-to-end AI project development.
Shipped ransomware detection on 500K+ hardware traces at sub-1% FPR. Built metric anomaly detection across 150+ time-series features and 32 KPI groups. Designed agentic LLM pipelines including an open-source MCP server and QLoRA-tuned diagnostic assistant. Productionized models on Docker, Kubernetes, and OpenShift AI. Led the 2024-25 AI Innovation Incubator.
Built multi-tenant telemetry pipelines ingesting 2 TB/day into PostgreSQL and Parquet on IBM Cloud. Streamed millions of metrics/hr via Kubernetes + Kafka. Cut small-file overhead by 40% with dynamic batching and size-aware flush logic.