I build systems
& chase ideas
worth testing.

Machine Learning Engineer Intern

Rancho Cordova, CA · Remote

rows through the pipeline

50M+rows through the pipeline

models shipped to prod

2models shipped to prod

AUC-ROC on 1:99 imbalance

0.85AUC-ROC on 1:99 imbalance

• Built an end-to-end pipeline over 50M+ rows — Snowflake ingest through model scoring to marketing-side export — and deployed 2 production models via Snowflake Model Registry on scheduled, YAML-configured runs.
• Implemented dynamic batched loading in SQL and Python for scheduled inference on 50M-row tables, and co-designed the deployment template and data contract that standardized the team's ML release process.
• Configured a Snowflake Cortex Agent that turns stakeholders' natural-language questions into optimized SQL, so they can generate their own graphs and tables.
• Built an Isolation Forest classifier on a 1:99 imbalanced dataset (0.85 AUC-ROC, 0.20 PR-AUC), plus a GAN synthetic-data pipeline that improved F1 on classification benchmarks.
• Hosted workshops on Gaussian Mixture Models, Transformers, GANs and Mixture of Experts for the Data Science department.

• Snowflake
• SQL
• Python
• Isolation Forest
• GANs
• MLOps

Quant Developer

University of Florida · Gainesville, FL

• Manage the fund's data engine on PostgreSQL, pgAdmin and Docker, orchestrating ETL pipelines for market-data ingestion and backtesting.
• Use Mixture-of-Experts and reinforcement learning to generate synthetic anomaly data, then train XGBoost + attention models to detect tail events.

• PostgreSQL
• Docker
• ETL
• XGBoost
• Reinforcement learning
• Mixture of Experts

AI Engineer Intern

Hanoi, Vietnam

alarms modeled

2.6Malarms modeled

fault classes

3.3kfault classes

held-out, Jamba-lite (vs 72% GRU)

85%held-out, Jamba-lite (vs 72% GRU)

• Trained an end-to-end PyTorch sequence model that predicts the next network fault from multi-domain alarm streams — 2.6M alarms, 126k sequences, 3.3k fault classes — selecting checkpoints by hyperparameter search on reproducible snapshots.
• Implemented a Jamba-lite hybrid (Mamba SSM + selective attention) against a GRU baseline — 85% vs 72% on the held-out set — and handled class imbalance with class weights, focal loss and label smoothing.
• Pushed single-GPU training to 80% utilization with TF32 mixed precision, pinned memory and tuned DataLoaders; reported Top-k accuracy and object-level recall dashboards.

• PyTorch
• Mamba / SSMs
• Sequence modeling
• TF32 / mixed precision
• Hyperparameter search

Machine Learning Research Assistant

University of Florida · Gainesville, FL

IoU on canopy segmentation

0.924IoU on canopy segmentation

less manual labeling

90%less manual labeling

ripeness prediction accuracy

+25%ripeness prediction accuracy

• Developed a CUDA-accelerated OpenCV pipeline integrating YOLO detection and SAM2 segmentation for video-scale plant tracking, reaching 90% reliability across full growth cycles and cutting manual labeling effort by 90%.
• Built a PyTorch LSTM that predicts ripeness from segmentation-derived fruit and temporal growth patterns, improving accuracy by 25%.

• PyTorch
• CUDA
• OpenCV
• YOLO
• SAM2
• LSTM
• Computer vision

FinTech Software Engineer Intern

Hanoi, Vietnam · On-site

param LLaMA 3.1 assistant

8Bparam LLaMA 3.1 assistant

over project documentation

RAGover project documentation

on-site internship

3 moon-site internship

• Developed an internal task-management assistant using LLaMA 3.1 8B integrated with Node.js to automate and streamline task handouts in a Waterfall development model.
• Implemented Retrieval-Augmented Generation (RAG) to query and synthesize project documentation for context-aware task assignments.
• Reduced manual coordination effort, improved clarity in task delegation, and built a scalable AI integration framework for structured workflows.

• LLaMA 3.1
• RAG
• Node.js
• LLMs

Project Captain, Launchpad Team

Gainesville, FL · Part-time

twitch-detection accuracy

97%twitch-detection accuracy

cost cut for small clinics

$3k+cost cut for small clinics

person team led

5person team led

• Take over stalled projects to finish development, document them thoroughly, and deploy deliverables for hospital and clinical use.
• Gator Goes to Surgery — led a five-person team building an interactive iOS app (Xcode, Cocos2d-X, C++) that eases children's anxiety before medical procedures; coordinated with Shands Hospital on App Store release and hospital integration.
• Train of Four — engineered an ML + MediaPipe twitch-response evaluation system at 97% accuracy that removes manual oversight and saves mid-to-small clinics over $3,000; deployed on Android.
• Milk Bank Optimizer — a calculator website that helps milk bank staff split donated milk into batches that meet calorie and protein targets.
• RAG chatbot — built a retrieval-augmented chatbot for Dream Team Engineering and the UF College of Medicine.

• C++
• iOS / Cocos2d-X
• MediaPipe
• Android
• RAG
• Leadership

Research Assistant

Hanoi, Vietnam · Part-time

• Completed an 8-week IoT and AI summer training program run by Hanoi University of Science and Technology with Makipos Electric Company.
• Supported senior students and professors in carrying out graduation projects and theses.

• C++
• IoT
• AI