S.P

Raleigh, North Carolina · fig. 01

Sakhi
Patel

I train models to see — satellite damage, genomes, live video.
Then I go home and paint.

Master’s student in Computer Science at NC State University. Machine learning, computer vision, and data engineering by training — a painter the whole time.

Watercolor painting of a glowing lantern hanging from a branch over a lavender field scattered with fireflies, against an orange dusk sky.
what the model sees / what I see — watercolor, from the paintings page
§ 01

Two instruments, one eye

I grew up in India and took my B.Tech in Information & Communication Technology at Pandit Deendayal Energy University. Now I’m in Raleigh for a Master of Computer Science at NC State, expected May 2027.

The thread through everything I make is looking carefully. A damage-classification model and a graphite portrait are trained the same way: you stare at the thing until you stop seeing what you expected and start seeing what’s there. One runs on a GPU, the other on a kitchen table — I don’t treat them as separate lives.

courseworkMachine Learning · Neural Networks · Database Management · Advanced Topics in Machine Learning

in progress, fall 2026Machine Learning with Graphs · Parallel Systems · Operating Systems

§ 02

Where I’ve worked

  1. May – Aug 2026

    Data Analyst Intern · KIOTI Tractor

    Consolidated dealer, sales, and territory data into a unified reporting layer with SQL Server stored procedures, and built Power BI dashboards tracking dealer status accuracy, territory alignment, and data completeness for Sales and Revenue Operations stakeholders.

  2. Jun – Jul 2024

    Data Analyst Intern · Bhavi Technologies

    Built an ETL pipeline integrating 5+ operational data sources, cutting processing latency 40% — real-time KPI reporting instead of weekly batches — plus Power BI and Tableau dashboards tracking 15+ KPIs, adopted by three business units within a month.

§ 03

Selected projects

F1 0.8136 one model vs. a 42-model ensemble (F1 0.8112)

Disaster Damage Detection

PyTorch · DeepLabV3+ · ResNet-50 · xBD benchmark · NC State, Spring 2026

A two-phase deep-learning pipeline that reads post-disaster satellite imagery: first localize every building, then classify how badly each one is damaged. I benchmarked seven geospatial foundation models for localization (the fine-tuned DeepLabV3+ won at Dice 0.877, beating Satlas and DINOv2), then designed JointDamageNet — a dual-branch network that beats the xView2 Challenge’s winning 42-model ensemble as a single model.

github.com/sakhi20/Disaster-Damage-Detection ↗
Six-panel model output for a Hurricane Michael tile: pre- and post-disaster satellite photos, ground-truth and predicted building masks, and ground-truth versus predicted damage maps colored by severity.
actual model output — hurricane michael tile, from the project report
18.9× compression — GZIP manages 3.2× on the same data

GenoCompress AI

TensorFlow · PyTorch · Conv1D autoencoder · NumPy · scikit-learn

Genomic datasets are enormous and repetitive — a good target for learned compression. This 1D-convolutional autoencoder compresses 500K+ DNA sequences 18.9× — six times better than GZIP — with accuracy validated at 92.3% using GATK benchmarking tools.

github.com/sakhi20/GenoCompress-AI ↗
ATGCGTACCGTTAGCATCGGATACCGTAGGCTAATCGCAGTACCGGTTAACGT…
schematic — sequences pass a conv1d bottleneck and come back 18.9× smaller
25 fps live inference · 87% precision · 0.85 mAP@0.5

Suspicious Activity Detection

YOLOv8 · OpenCV · undergraduate capstone, PDEU 2024

Led a three-person team building a real-time surveillance system that detects weapons and masks in live video, with a proximity-based filter that cut false positives in crowded scenes by 20%. It holds 87% precision and 83% recall at a sustained 25 frames per second.

github.com/sakhi20/AI-Powered-Suspicious-Activity… ↗
A live CCTV frame from the system: a person in an alley outlined by a blue detection box labeled Suspicious.
live detection frame — from the capstone report
§ 04

Toolbox

Languages
Python (Pandas, NumPy, scikit-learn) · SQL (MySQL, PostgreSQL, SQL Server) · R
ML & vision
PyTorch · TensorFlow · XGBoost · OpenCV · YOLOv8 · NLP
Viz & BI
Power BI · Tableau · Matplotlib · Seaborn
Data engineering
ETL pipeline design · data warehouse development · SSMS · SharePoint
Tools
Git / GitHub · Docker · AWS · Linux · Jupyter · VS Code
Off-hours
Acrylic · watercolor · graphite — see the paintings

§ 05 · correspondence

Say hello.

Looking for research positions, internships, and co-ops in machine learning and data science. Or just talk paintings — that works too.