Nagaraj Loni

Computer Science graduate - Machine Learning & AI Engineering

I build production-ready ML systems: pipelines that train, track, containerize and ship models - not just notebooks that prove a point.

linkedin /in/nagaraj-loni github nsgowebjavaprog leetcode NS258 gmail nagarajloni209@gmail.com mobile +91 63615 44673
72%ROC-AUC · churn model
20K+rows scored / request
9.29CGPA · out of 10
6certifications & awards

About

Introduction

I'm a Computer Science Engineering graduate fascinated by the gap between a model that scores well and one that people can actually rely on.

That's the thread through my work - KannadaLM, a monolingual LLM I trained from scratch, taught me what happens under the hood of a language model; my churn-prediction and vehicle-insurance platforms taught me how to wrap that model in the plumbing that makes it usable - FastAPI services, MLflow tracking, Docker containers, and CI/CD that ships without babysitting.

I'm comfortable across the stack: Python and ML on one end, React and REST APIs on the other, AWS and Docker holding it together in between. What I'm looking for next is a team building real ML products, where I can keep closing that gap between prototype and production.

Education

Education

B.E - Computer Science & Engineering

Ballari Institute of Technology and Management (BITM), Ballari, Karnataka

2022 – 2026 CGPA 9.29 / 10

2nd PUC

BIPU Badagandi, Karnataka

2020 – 2022 Percentage 86.67%

Skills

Skills

Languages & Databases

  • Python
  • MySQL

AI / ML & Data Science

  • Machine Learning
  • Deep Learning
  • NLP
  • Generative AI
  • Large Language Models (LLM)
  • RAG
  • Feature Engineering
  • PyTorch
  • LangChain
  • NumPy
  • Pandas
  • Matplotlib
  • Scikit-learn

MLOps & Cloud

  • Docker
  • MLflow
  • DVC
  • CI/CD Pipelines
  • Git
  • FastAPI
  • REST APIs
  • AWS (EC2, S3)

CS Fundamentals

  • Data Structures & Algorithms
  • Operating Systems
  • OOP

Projects

Projects

churn_prediction_platform.py

End-to-End Customer Churn Prediction Platform

view repo →

Aug 2026

A full MLOps pipeline that benchmarks Logistic Regression against Random Forest, auto-selects the better model from 10+ tracked experiments, and serves predictions through a production-style API and UI.

  • Built the pipeline with NumPy, Pandas & Scikit-learn, reaching 72% ROC-AUC and tracking every run in MLflow to auto-select the best model - a ~15% recall lift over baseline.
  • Designed a FastAPI backend (Pydantic, 15+ REST endpoints, full CRUD on SQLAlchemy/SQLite) with a batch CSV pipeline that validates and scores 20,000+ rows per request.
  • Paired the backend with a React.js frontend for real-time predictions, upload, and history.
  • Containerized the stack with Docker & Docker Compose (multi-stage builds), added Pytest coverage for validation/CRUD flows, and managed the codebase with Git.
PythonScikit-learnMLflowFastAPIReactDockerPytest
vehicle_insurance_platform.py

Vehicle Insurance Prediction Platform

view repo →

Nov 2025 – Dec 2025

An end-to-end machine learning platform for vehicle insurance risk prediction, built for reliable, repeatable retraining rather than a one-off model.

  • Built the ML core with Python, NumPy & Pandas, containerized with Docker and orchestrated with Kubernetes.
  • Implemented a complete MLOps pipeline - Git, DVC, Docker, FastAPI, MLflow and AWS S3 for model versioning, training, testing and automated CI/CD deployment.
  • Achieved production-grade reliability through automated model retraining and deployment workflows.
PythonPandasDockerKubernetesDVCMLflowAWS S3

Achievements

Achievements & Certifications

Six markers of the same habit - finishing what I start. Click any card to open the certificate.

★

Also received the "Best Tech Stack User" award at a national-level hackathon.

Experience

Experience

Jan 2026 Mar 2026

AWS Cloud Computing - Virtual Internship

SmartBridge, in partnership with NASSCOM FutureSkills Prime

  • Completed a structured virtual internship covering core AWS cloud services and architecture concepts.
  • Applied AWS concepts hands-on, building practical familiarity with cloud deployment models and services (S3, EC2).
  • Deployed the Vehicle Insurance Prediction Platform using S3 for automated CI/CD model deployment.

Q&A

Q&A

Click one to expand the answer.

I'm a Computer Science Engineering graduate from BITM Ballari with a 9.29 CGPA, focused on machine learning and AI engineering. I trained KannadaLM, a monolingual LLM, from scratch, and built two end-to-end MLOps platforms - a customer churn predictor and a vehicle insurance risk model - using FastAPI, MLflow, Docker and AWS. I care less about a model that scores well in a notebook and more about one that survives contact with real traffic, which is why I lean into the pipeline and deployment side as much as the modeling itself.

Strength: I finish the unglamorous parts. It's easy to get a model to 70% accuracy in an afternoon; I'm the person who then adds the CI/CD pipeline, the test coverage and the containerization so it actually ships - that's how both my major projects ended up production-shaped, not just prototype-shaped.

Weakness: I can over-invest in infrastructure before validating the idea is worth it. I've gotten better at time-boxing the "make it robust" phase and asking whether the model is even the right one before I containerize it.

Because I already work like an engineer, not just a modeler. My churn platform has 15+ REST endpoints, full CRUD, Pytest coverage and Docker Compose orchestration behind it - that's the same shape of work a production ML team does daily. I'm AWS Certified, comfortable with Docker/Kubernetes/MLflow/DVC, and I pick things up fast: KannadaLM meant learning to train a language model from first principles in a short window.

Owning ML systems end-to-end - not just building a model, but being the person a team trusts to take it from experiment to a reliable service in production, and mentoring others through that same jump. I'd like to keep going deeper into MLOps and LLM infrastructure specifically, since that's where I already spend most of my time by choice.

Training KannadaLM from scratch was the hardest thing I've built - there was no fine-tuning shortcut, I had to get the data pipeline, tokenization and training loop right with limited compute. I broke it into stages, tracked every experiment so I could tell what actually moved the needle, and treated failed runs as data rather than setbacks. That same "instrument everything, iterate in small steps" habit is what later made MLflow tracking and DVC feel natural on my other projects.

Connect with Me.

Let's build something reliable.

Open to ML engineering, MLOps and applied AI roles. If a system needs to go from a working model to something a team can depend on, I'd like to hear about it.

Email me Connect on LinkedIn See my code