About
The best software is smart. The best smart software is boring, in the sense that it just works.
I'm a Computer Science student at UC Irvine, currently an MTS Intern at Nutanix and an AI/ML Fellow with Break Through Tech, joining AWS as an SDE Intern in the fall. My work covers AI powered infrastructure, agentic workflows, computer vision, and the backend systems.
I'm especially interested in taking a model or an agent from a promising experiment to something that runs in production, survives bad inputs, and quietly does its job every day.
Open to connect! Chat with me about AI, infrastructure, or anything in between.
Experience
Amazon Web Services (AWS)
Software Development Engineer Intern
- Incoming Fall 2026.
Nutanix
Intern, Member of Technical Staff
- On Neuron team: The core Intelligent Operations (AIOps) service for Nutanix Cloud Manager (NCM), driving enterprise-scale analytics services for hyperconverged infrastructure (HCI).
- Built agentic workflows for bug triaging and root case analysis.
Break Through Tech
AI/ML Fellow
- Year-long fellowship by Cornell Tech; aims to bridge the gap between a technical degree and an industry-ready AI engineer. Selected into a national cohort spanning a network of university and industry partners.
- Completed full summer course on ML designed by Cornell. In the process of delivering an industry ML challenge project scoped by industry sponsors.
Bettascore.ai
Software Engineer, AI Intern
- A risk control system for electric scooter rentals, on AWS: fraud detection, credit scoring, and order validation as rides happen.
- A FastAPI service in front of PostgreSQL, with a decorator driven dispatcher routing each order to its scoring module.
- Redis cache aside with TTL expiration. Hit rate sits at 89%, which takes about 75% of the load off the database at peak.
- Lambda and SQS chain into a ten stage risk pipeline, with adversarial approval gates between stages.
- Prometheus watches API latency, fraud detection rates, and the rest, so drift shows up on a dashboard first.
Boundary Remote Subsurface Solutions
Machine Learning Engineer (Contract)
- GeoSAE in PyTorch: stacked autoencoders that turn sparse borehole logs into 3D stratigraphic models. Sequence ordering, dip and strike, and an Eikonal smoothness term live in the loss itself, so the geology constrains training rather than getting corrected after it.
- Sparse convolutional autoencoders for interpretability on resistivity grids, over formations pulled from the Macrostrat API. Sampling patterns follow how sparsely a survey actually gets drilled, which keeps validation honest about field conditions.
Beckman Laser Institute
Machine Learning Engineer Intern
- A cell segmentation pipeline for microscopy datasets. Analysis that took 2+ hours by hand now finishes in under 5 minutes.
- Preprocessing that copes with 500GB+ of microscopy data: quality filtering, normalization, tiling.
- A Cellpose model fine tuned in PyTorch on 1,200+ images. 92% Dice, 15% over baseline.
- Experiments tracked in MLflow across CNN restoration configs, which is how the 25%+ PSNR gain was found.
UCI Health, Gavin Herbert Eye Institute
Undergraduate Researcher
- Undergraduate researcher at Lin Lab.
- A Bayesian decision analytic model that weighs surgical strategies for glaucoma so the choice rests on numbers. Co-authored as “A Quantitative Decision Framework for Surgical Intervention in Normal Tension Glaucoma using Bayesian Inference and Decision Theory,” under revision at the Journal of Ophthalmic and Vision Research.
UC Irvine, Donald Bren School of ICS
Learning Assistant
- Lab sections and office hours for the ICS 31/32/33 intro sequence. Most of the job is reading someone else's code closely enough to find where the reasoning broke, not where the syntax did.
- Grading and feedback on assignments, plus revisions to the course materials. A misconception that keeps turning up in office hours is usually a problem with the handout, not the student.
Projects
Three systems built end to end: a document Q&A agent, a campus carpooling platform, and a fault tolerant job scheduler.
AgentFlow
LLM Powered Document Q&A Agent
- Python
- LangChain
- LangGraph
- FastAPI
- PostgreSQL
- Redis
- AWS
- Docker
- 01A LangGraph agent that answers questions over documents and runs workflows through LangChain tools.
- 02pgvector for embeddings, Redis for repeat answers. About 45% fewer LLM calls.
- 03Docker throughout, GitHub Actions for CI/CD, with evaluation gates ahead of each deploy.
- 01A carpooling platform for UCI students, in React and TypeScript. An LLM handles the matching, so riders get paired on more than pickup points.
- 02Routes drawn with Mapbox GL geocoding and directions. Supabase PostgreSQL behind it, with Row Level Security keeping each user's data their own.
- 03Serverless end to end: Realtime push notifications, CI/CD, and React.lazy splitting to keep first load light.
Distributed Task Scheduler
Fault Tolerant Job Orchestration
- Go
- Redis
- PostgreSQL
- Docker
- Kubernetes
- Prometheus
- 01A distributed scheduler in Go. Redis priority queues feed worker pools that hold 50K+ concurrent jobs.
- 02Job metadata, history, and results land in PostgreSQL, so there is always an audit trail to read back.
- 03Docker and Kubernetes underneath, with workers autoscaling on queue depth.
Stack
from the languages up through the systems that keep a service running.
Frames
Some of my favorite photos.
UCI Middle Earth Dome of Stars
7:00 PM, 9/27/2024
Santa Fe Dam Recreational Area
7:00 PM, 8/13/2021
Middle Earth Shadowfax @ UCI
4:11 PM, 3/1/2024
Aldrich Park @ UCI
2:47 PM, 4/15/2024
Brandywine Bridge @ UCI
7:54 PM, 4/18/2024
Contact
Chat with me about AI, infrastructure, or anything in between.
Let’s build something boring.