Full Stack Engineer at Fidelity Investments

Hi, I'm Rhea Bhatia

Building scalable applications, real-time data pipelines, and machine learning systems that transform complex data into actionable insights.

View My Work Contact Me
GitHub LinkedIn

About Me

I'm a Full Stack Engineer at Fidelity Investments, where I build and optimize a high-performance, real-time trading application used by thousands of users.

I enjoy working across the stack, from designing backend systems and APIs to building user-facing interfaces, with a focus on performance, scalability, and reliability.

Outside of work, I'm deeply interested in data science, machine learning, and data engineering. I enjoy building projects that explore data analysis, model development, and real-time data systems, especially when they involve turning large volumes of raw data into meaningful insights.

Areas of Interest

Full Stack Systems

(Backend-focused)

Data Science & Machine Learning

Data Engineering & Real-Time Pipelines

AI-Driven Applications

Featured Projects

Deep dives into building scalable data engineering pipelines and machine learning systems from scratch to production.

Machine Learning Data Science Model Deployment Backend Systems

Predictive Churn Engine

End-to-end data science and machine learning system for the telecom industry with API deployment and monitoring using XGBoost and FastAPI.

View Full Breakdown
Data Engineering Distributed Systems Real-Time Systems Backend Systems

AI-Powered Real-Time Log Monitoring System

Real-time data pipeline and AI-powered system for log ingestion, anomaly detection, and incident analysis using Kafka, ClickHouse, and LLaMA 70B.

View Full Breakdown

Let's Connect

Drop me an email Connect on LinkedIn
Back to Projects
Machine Learning Data Science Model Deployment Backend Systems

Predictive Churn Engine

End-to-end data science and machine learning system with API deployment and monitoring

Python Pandas NumPy Scikit-learn XGBoost FastAPI Docker Pytest

🚨 Problem

In the telecom industry, customer churn directly impacts revenue, but identifying at-risk customers is challenging due to the complex interplay of behavioral and financial factors, including contract type, tenure, service usage, and billing patterns.

Without a data-driven system, telecom providers rely on reactive strategies, making it difficult to detect early warning signals of churn, resulting in lost revenue and missed opportunities for targeted customer retention.

🚀 Key Highlights

  • ✓ Conducted exploratory data analysis (EDA) to identify key churn drivers
  • ✓ Engineered features to better capture customer behavior patterns
  • ✓ Built and optimized XGBoost model (ROC-AUC: 0.85)
  • ✓ Tuned decision threshold (0.33) for recall-focused predictions
  • ✓ Deployed model via FastAPI with drift detection and automated testing

🔍 Full Breakdown

📊 Data Exploration & Insights

  • Analyzed customer behavior across tenure, contract type, and billing
  • Identified high-risk churn segments such as month-to-month contracts, short tenure customers, and high monthly charges
  • Validated insights through visual EDA and statistical analysis

🧬 Feature Engineering

  • Created TotalChargesPerMonth to capture spending behavior over time
  • Cleaned dataset by handling missing and non-numeric values
  • Applied preprocessing pipeline with one-hot encoding and feature scaling

🤖 Modeling & Evaluation

  • Trained Logistic Regression (baseline) and XGBoost (final model)
  • Performed hyperparameter tuning using RandomizedSearchCV
  • Achieved ROC-AUC of 0.85, accuracy of 78.1%, and F1 score of 0.65

⚖️ Business Optimization

  • Tuned classification threshold to 0.33 to prioritize recall
  • Focused on identifying high-risk customers for retention strategies

🚀 Production System

  • Built FastAPI service with endpoints /health, /model-info, and /predict
  • Implemented input validation using Pydantic
  • Designed API for seamless frontend or service integration
  • Added drift detection script for monitoring model performance
  • Containerized with Docker and added pytest-based tests

📈 Impact

  • ★ Enables proactive churn prediction using data-driven insights
  • ★ Demonstrates full lifecycle from data exploration to deployment and monitoring
  • ★ Bridges data science and backend engineering in a production-ready system
View on GitHub
Back to Projects
Data Engineering Distributed Systems Real-Time Systems Backend Systems

AI-Powered Real-Time Log Monitoring System

Real-time data pipeline and AI-powered system for log ingestion, anomaly detection, and incident analysis

Kafka / Redpanda ClickHouse Python Streamlit Groq (LLaMA 70B) Docker

🚨 Problem

Modern distributed systems (e.g., microservices architectures and trading platforms) generate millions of logs per minute, making it extremely difficult to identify meaningful patterns and detect critical issues in real time.

Failures such as API timeouts, database outages, and infrastructure bottlenecks are often buried in noisy log streams, leading to delayed incident detection and inefficient debugging workflows.

🚀 Key Highlights

  • ✓ Built a real-time log ingestion pipeline using Kafka and ClickHouse
  • ✓ Designed stream processing workflow for continuous log analysis
  • ✓ Implemented anomaly detection using time-based error spike patterns
  • ✓ Integrated LLM for automated root cause and incident report generation
  • ✓ Built interactive frontend dashboard using Streamlit for real-time visualization

🔍 Full Breakdown

📊 Data Exploration & Insights

  • Analyzed log distributions across severity levels (INFO, WARN, ERROR)
  • Identified error spikes in short time windows as key anomaly indicators
  • Designed aggregation queries to surface high-frequency system failures

🧬 Feature Engineering

  • Structured logs into queryable schema (event_id, level, message, timestamp)
  • Created materialized views for fast aggregation in ClickHouse
  • Engineered time-based features (e.g., errors per minute) for anomaly detection

🤖 Modeling & Analysis

  • Implemented rule-based anomaly detection using statistical thresholds
  • Used LLM (LLaMA 70B) for root cause analysis, correlation detection, and impact assessment

⚖️ System Optimization

  • Triggered AI analysis only during anomalies to reduce cost and latency
  • Prioritized real-time responsiveness and efficient querying

🚀 Production System

  • Designed distributed pipeline: Producer → Kafka → Consumer → ClickHouse
  • Built real-time dashboard using Streamlit
  • Containerized services using Docker

📈 Impact

  • ★ Enables faster detection of system failures through real-time analytics
  • ★ Reduces manual debugging using automated AI-driven insights
  • ★ Demonstrates production-style data engineering pipelines, backend systems, and AI integration
View on GitHub