M.S. Computer Science · Los Angeles, CA

Building AI systems that ship: agentic LLM pipelines, healthcare ML, and the engineering rigor in between.

I'm a recent M.S. Computer Science graduate from USC (May 2026) with a background in software and machine learning engineering. I've worked across banks, fintech, and AI startups, and I like problems that sit at the intersection of well-designed systems and real machine learning, where the work has to be correct, fast, and shippable.

Right now I'm focused on agentic LLM systems, applied NLP, and ML for healthcare. I care a lot about the production side: containerized inference, CI/CD, fine-tuning that actually moves a metric, and pipelines that don't fall over at 1,200 requests an hour.

🔭 Currently: contributing to open-source ML libraries and fine-tuning an open model to publish on Hugging Face Hub.

Outside of code, I think about sustainability, interdisciplinary AI, and the bigger questions about why any of this matters.

Languages

Machine Learning & AI

Data & Streaming

Cloud & DevOps

Monitoring & Observability

A production-style incident response orchestrator built on a 10-node Temporal DAG, with Claude-powered root cause analysis driving the diagnostic path. Tested across 230+ synthetic incidents, with MTTR dropping from 47 minutes to 18 minutes. TypeScript · Next.js · Temporal · Claude API · Railway

A FastAPI service for radiology prior relevance prediction, deployed on Render. TF-IDF with n-gram features feeding a scikit-learn classifier, hitting 98.27% accuracy on the public smoke-test set. Designed to slot into a healthcare ML workflow as a low-latency relevance scorer. Python · FastAPI · scikit-learn · Render

Fine-tuned a 4-bit quantized LLaMA 3.2 Vision model on the MultiUI/GUI dataset using Unsloth, exploring efficient multimodal adaptation under tight memory budgets. PyTorch · Unsloth · LoRA · LLaMA 3.2

Fine-tuned distilbert-base-uncased on the dair-ai/emotion dataset (6 emotions) for multi-class emotion detection, reaching 94.25% accuracy and 0.9171 macro-F1 on the eval set, and published to the Hugging Face Hub. Followed up with a DeBERTa-v3-small experiment using class-weighted loss and a clean ablation to lift macro-F1 on the rare emotions. PyTorch · Transformers · DistilBERT · DeBERTa-v3 · Hugging Face Hub

🥇 1st Place, EcoMate AI Hackathon (Sustainable Infrastructure category) A Streamlit app that estimates personal carbon footprints by extracting activity data from natural-language text and images using a multimodal GenAI pipeline. Python · Streamlit · GenAI · Multimodal

A custom YCSB Java binding for JanusGraph, with schema and CRUD driven through remote Gremlin traversal. Used to analyze single-node vs. multi-node performance characteristics under standard workload mixes. Java · JanusGraph · Gremlin · YCSB

Merged PRs shipping in widely used Python & ML libraries (millions of downloads a month):

• huggingface_hub: 4 merged PRs, 3 shipped in the v1.22.0 release and credited in the changelog. Made filter_repo_objects pattern matching case-sensitive across all platforms (#4435), extended parse_size to accept two-letter byte units (KB/MB/GB/TB) (#4468), and corrected the http_backoff retry behavior plus CLI help (#4436, #4477)
• networkx: aligned boruvka_mst_edges defaults with Kruskal/Prim (#8728)
• dpgen: fixed an AttributeError in ABACUS Gamma post-processing (#1920)
• xarray: docstring fix in cumulative() (#11425)
• statsmodels: docstring fixes (#9873)

All merged PRs →

• Software Engineer Intern at The Verse (May - Aug 2025): Agentic LLM pipelines across 47 daily workflows, BART entity extractor fine-tuned from 0.74 to 0.91 F1, containerized inference handling 1,247 req/hr at peak.
• Software Engineer at Bank of America: Production backend systems and data workflows; automated Kafka topic & role provisioning (Enterprise-Level Silver Award).
• Software Engineering Intern at HighRadius: AI-Enabled FinTech B2B invoice management using XGBoost.

• 🥈 Enterprise-Level Silver Award, Bank of America: for automating Kafka topic & role provisioning, replacing a manual multi-team workflow.
• 🥇 1st Place, EcoMate AI Hackathon: Sustainable Infrastructure category, for EcoMate-AI (live demo), a multimodal GenAI carbon-footprint estimator.
• 🔀 Shipped open-source contributor: code merged into Hugging Face Hub, NetworkX, xarray, statsmodels, and dpgen, with 3 PRs live in the Hugging Face Hub v1.22.0 release (details above).

• Long-standing interest in reinforcement learning, evolutionary optimization, and applied LLM research.
• Thinking a lot about AI for healthcare and education, plus sustainability as a design constraint, not an afterthought.
• Curious about the deeper why behind the work: spirituality, meaning, and what good engineering owes the people it touches.

Open to opportunities in ML engineering, software engineering, NLP, and healthcare AI.