I'm a first-year MS Computer Science student at Vanderbilt University and a Research Assistant at the VU-BEAM Lab, where I work on self-supervised medical imaging using contrastive learning. My work sits at the intersection of research and production — I care about both the theory and actually shipping things.
My current research focus is DAC-Learn (Depth-Aware Contrastive Learning), a novel self-supervised beamforming method for ultrasound imaging evaluated on the PICMUS benchmark. Outside the lab, I build production-grade systems — an API gateway handling 25k+ req/s, a full crisis operating platform on PostGIS, and an AI agent evaluation environment.
Open to Summer 2026 internships — SDE, ML Engineering, and Research roles.
| DAC-Learn | Depth-Aware Contrastive Learning for ultrasound beamforming. Novel self-supervised method evaluated on PICMUS benchmark against SimCLR, BYOL, CycleGAN, and supervised baselines. Ongoing — targeting a top-tier ML venue. |
| VU-BEAM Lab | 85% improvement in image contrast (CNR: 3.5+ dB) using SimCLR with custom NT-Xent loss on 10,000+ unlabeled cardiac ultrasound images. Outperforms CycleGAN baselines by 20% on held-out clinical data. |
|
Full-stack crisis platform built around a shared live PostGIS graph. Citizens get safe routing via modified Dijkstra with time-decaying edge weights. Responders get crew sequencing, XGBoost ML-predicted circuit failures, and Claude-powered NLP triage — all synchronized via Socket.io in real time.
|
Production-grade Node.js gateway achieving 25,000+ req/s at P99 < 15ms. Google SRE error-budget model, distributed tracing with waterfall flame-graph, EWMA PID adaptive rate limiter, circuit-breaker state machine, and 10-tab live admin dashboard. Deployed on AWS ECS.
|
|
Multi-model NLP pipeline across 7 fine-tuned transformers (TinyLLaMA, Mistral 7B, ProphetNet) with LoRA fine-tuning, instruction tuning, and W&B experiment tracking. Achieved 15% ROUGE/BLEU improvement over baselines across 200+ documents. Awarded Best Paper at IEEE international conference.
|
Realistic benchmark environment for evaluating AI agents on multi-step business data-cleaning workflows. Agent interacts with messy CRM, orders, and payments tables via typed env APIs and receives deterministic scores with reward shaping for partial credit. Includes task design, grading logic, tests, and a deployable server.
|
|
Fine-tuned Code Llama 7B with QLoRA on a custom 10,000-sample dataset covering OWASP API Top 10 across 10+ languages. 3-stage pipeline: structural endpoint discovery → LLM inspection → fuzzy OpenAPI policy validation. Streamlit dashboard + CLI.
|
React Native + Node.js mutual-aid dispatch platform. Intake flow, volunteer matching, OTP-verified handoff, charity fallback, and Socket.io real-time sync across mobile clients. JWT-secured REST API with session-based auth.
|
Languages
AI / ML
Backend & Infrastructure
- 🏆 IEEE Best Paper Award — Paper2Story: Multi-Model Narrative Generation Pipeline (2024)
- 📄 IEEE Published — Hybrid API Vulnerability Detection System (2025)
- 🔬 Ongoing Research — DAC-Learn: Depth-Aware Contrastive Learning for Ultrasound Beamforming (targeting top-tier ML venue)
- 🔬 Developing DAC-Learn — novel depth-aware contrastive learning for ultrasound beamforming at VU-BEAM Lab
- 🏗️ Building a distributed task queue and job-matching RAG pipeline
- 📖 Grinding NeetCode 150 daily for SDE interviews
- 🤝 Open to Summer 2026 internships — SDE, ML Engineering, Research
