Open to ML/AI engineering work that ships

Sai Nikhil Chidipothu.

ML/AI Engineer

Master's student in Computer Science at University of Central Florida, Graduate Teaching Assistant for Algorithms for Machine Learning and Computer Graphics. Building applied ML, computer vision, and LLM efficiency work.

Explore work Résumé Contact

01 / About

A little about me

I like building things that start as a rough idea and slowly become something clear, testable, and useful. Most of my work sits around machine learning, computer vision, and efficient inference, but I care just as much about the parts around the model.

My default style is simple: get a working baseline, understand what is actually improving, and keep the code readable enough that I can come back to it a month later without hating myself.

In practice

  • I start simple before making anything clever.
  • I like projects with a clear metric, constraint, or user.
  • I write code for the next person reading it, even when that person is me.
  • I enjoy turning ML experiments into something people can actually try.

Good overlap

Applied ML, computer vision, LLM efficiency, and the engineering work that makes models easier to run, evaluate, and use.

02 / Selected work

Selected work with context

A few projects that show how I think: define the problem, make tradeoffs visible, measure the result, and build something someone could actually use.

Project 01

SparseGPT · One-Shot LLM Pruning

Completed · LLM efficiency

Made LLaMA-7B lighter without retraining, then measured the real cost of that speedup instead of assuming compression is free.

  • Built for: efficient inference experiments where GPU memory, latency, and model quality all matter.
  • My work: implemented SparseGPT runs across unstructured and 2:4 / 4:8 structured sparsity.
  • Measured: perplexity and zero-shot accuracy on ARC-C, ARC-E, and PIQA at 25 to 60% sparsity.
  • Outcome: completed pruning benchmarks across sparsity settings and documented the accuracy/compression tradeoffs.
  • PyTorch
  • Hugging Face
  • CUDA
  • Quantization
Code

Project 02

SkyEye · Aerial Object Detection

Computer vision · real-time inference

A drone-vision pipeline for turning aerial footage into fast, usable detections rather than offline-only model outputs.

  • Built for: scenarios where operators need quick object detection from aerial frames.
  • My work: led YOLOv8 fine-tuning, inference tuning, and the model evaluation loop.
  • Outcome: reached 93% detection accuracy on 5,000+ annotated drone frames.
  • Runtime: kept inference under 80 ms per frame and wrapped the pipeline in Streamlit.
  • YOLOv8
  • PyTorch
  • OpenCV
  • Streamlit
Code

Project 03

What-TO-DO · Productivity Tracker

Live product · actively maintained

A productivity app I actually use to keep classes, tasks, goals, and daily notes in one place.

  • Built for: students balancing coursework, project work, schedules, and daily planning.
  • My work: designed and shipped the React + Vite app with Tailwind and Vercel deployment.
  • Engineering: structured the app with reusable components, hooks, and utilities so new features stay easy to add.
  • Product detail: added semester-scoped class schedules so planning matches how school actually works.
  • React
  • Vite
  • Tailwind CSS
  • Vercel
Live demo Code

Project 04

Pill Detection and Identification

Published · JETIR, May 2024

A medical-image classification project focused on reducing visual confusion between similar-looking pills.

  • Built for: safer pill identification from uploaded images, especially when pills look visually similar.
  • My work: trained a MobileNetV2 transfer-learning classifier on 10,000+ pill images.
  • Outcome: achieved 95% classification accuracy and published the work in JETIR, May 2024.
  • Product layer: deployed a Flask app with Bootstrap UI and Matplotlib result visualizations.
  • TensorFlow
  • Keras
  • MobileNetV2
  • Flask
  • Bootstrap
Code

Project 05

Phishing Website Detection

Security ML · model comparison

A security-focused ML benchmark for spotting phishing URLs while paying attention to false positives.

  • Built for: browser or email-security workflows where a bad false positive can break trust.
  • My work: engineered lexical and structural URL features from the PhishTank dataset.
  • Compared: Random Forest, SVM, and Gradient Boosted Trees across precision, recall, and false positive rate.
  • Takeaway: treated model selection as a product tradeoff, not just a leaderboard score.
  • Python
  • Scikit-learn
  • Pandas
  • Flask
Code

03 / Toolkit

What I work with

Grouped the way I'd actually pick them up on a project, not the way an ATS spits them out.

Languages

  • Python
  • Java
  • C
  • SQL
  • JavaScript
  • HTML
  • CSS

ML & AI

  • PyTorch
  • TensorFlow
  • Keras
  • Scikit-learn
  • OpenCV
  • YOLOv8
  • MobileNetV2
  • Hugging Face
  • Transformers
  • LangChain
  • RAG
  • AI Agents
  • NLP

Web & APIs

  • FastAPI
  • Flask
  • REST
  • React
  • Tailwind CSS
  • Bootstrap
  • Streamlit
  • Postman

Cloud & DevOps

  • AWS EC2
  • AWS S3
  • AWS RDS
  • AWS Lambda
  • Docker
  • Kubernetes
  • Kafka
  • Airflow
  • GitHub Actions
  • Jenkins
  • CI/CD
  • Linux

Data

  • PostgreSQL
  • MongoDB
  • PySpark
  • Pandas
  • NumPy
  • Power BI
  • EDA
  • Feature Engineering

Graphics

  • OpenGL
  • ModernGL
  • Pygame

Relevant graduate coursework

Algorithms for Machine Learning · Advanced AI · Computer Vision · Computer Graphics · Computer Forensics · Foundations of Computer Security and Privacy · Current Topics in Machine Learning

04 / Contact

Contact

Email is the simplest path. I’m also active on LinkedIn and GitHub if that is easier.

Best way to contact me

For opportunities, research conversations, or project questions, a short email with context is perfect.

LinkedIn GitHub Résumé