MLOps Projects | Amulya Gupta — Production ML Pipelines & AI Systems

Project 01

Heart Disease Prediction — End-to-End MLOps System

MLOps Docker Kubernetes FastAPI MLflow Prometheus Grafana GitHub Actions Python

Problem Statement

Heart disease prediction models are commonly trained in isolation — but rarely deployed reliably. The challenge: build a full MLOps lifecycle, not just a model, that can be reproduced, versioned, deployed, and monitored in a real infrastructure environment.

What I Built — Architecture

Data & Training Layer: Preprocessing pipeline with reproducible artifact management using MLflow experiment tracking

Inference Layer: FastAPI service with input validation (Pydantic), health check endpoints, and Prometheus metrics exposure

Infrastructure Layer: Docker containerisation + Kubernetes (Minikube) orchestration with declarative manifests

Observability Layer: Prometheus scraping metrics + Grafana dashboards for inference latency, error rates, and request volume

CI/CD Layer: GitHub Actions pipeline automating build, test, and deployment on every push

Key Technical Decisions

▸ Chose MLflow over manual artifact tracking for full reproducibility — any experiment can be re-run identically
▸ FastAPI over Flask for async support, built-in OpenAPI docs, and Pydantic validation — production-appropriate choice
▸ Kubernetes even on Minikube to prove the manifests work in a real orchestration environment

Outcome & Learnings

Delivered a fully containerised, observable ML inference system with automated deployment — demonstrating the complete MLOps lifecycle from raw data to production monitoring. Key learning: observability is not optional; you need to know your model is behaving correctly in production.

Project 02

Customer Churn Prediction — Automated ELT Pipeline & ML System

View on GitHub ↗

Python Prefect ETL Pipeline Supervised Learning Data Engineering Automation PostgreSQL

Problem Statement

Customer churn data typically comes from multiple sources — CRM, usage logs, billing — and is messy. Building a reliable, automated pipeline to ingest, clean, consolidate, and feed this data into a model training loop is the real engineering challenge.

What I Built

ELT Pipeline: Python-based orchestration using Prefect for multi-source ingestion, transformation, and consolidation

Data Quality Layer: Cleaning and validation steps ensuring training data consistency

Model Layer: Trained and evaluated multiple supervised learning models, selecting highest performer at 90%+ accuracy

Performance Optimization: Reduced pipeline execution time by ~40%, enabling dramatically faster retraining and experimentation cycles

Outcome

A production-ready automated ML pipeline that processes multi-source customer data, trains and evaluates models, and can be re-triggered on schedule — 40% faster than the initial baseline implementation.

Project 03 — Production Case Study

Adobe Journey Optimizer — Production Campaign Platform at Scale

ℹ️ This is a case study of my production work at HCLTech. No proprietary code is shared — only outcomes, decisions, and learnings.

Production Adobe Journey Optimizer 20M+ Users Campaign Engineering Data Pipelines Personalization

What I Owned

▸ End-to-end ownership of large-scale in-app campaign delivery using Adobe Journey Optimizer
▸ Delivered 40+ production campaigns during peak Black Friday/Cyber Monday traffic periods
▸ Reached 20M+ users while improving CTR and purchase completion by 10–15%
▸ Owned production readiness: QA coordination, experience rendering validation, stable releases
▸ Built and optimized data ingestion workflows supporting personalised decisioning systems

Engineering Highlights

Zero critical post-launch issues across 40+ high-stakes campaigns during the most traffic-intensive retail periods of the year. This requires meticulous production readiness, audience segmentation discipline, and cross-functional coordination across engineering, QA, and product.