Rohan Murugan Dashboard

Data scientist · Dallas, TX · open to relocation

Rohan Murugan

Machine learning for fraud detection, risk and customer behavior

I build ML models and data pipelines that lift accuracy by 20%, cut latency by 25–30% and keep 1M+ records clean enough to trust, then explain the results to the people who act on them.

Key metrics

Aptech · model accuracy

+20%

higher accuracy from PySpark and SQL models on 1M+ record healthcare datasets

Aptech · data integrity

99.5%

data integrity ensured across datasets of more than one million records

Aptech · latency

−25%

lower latency for 50,000 users after ETL pipelines and REST microservices

Sinebyte · chatbot

−30%

lower response latency on a GPT-based chatbot built with Flask and ReactJS

Production pipeline at Aptech

1M+ records to 50,000 users
  1. Ingest1M+ recordshealthcare datasets, 99.5% integrity
  2. TransformPySpark + SQLETL pipelines
  3. ModelML models+20% accuracy
  4. ServeREST APIsmicroservices in production
  5. Users50,000 users25% lower latency

Simplified architecture of the work described in the first case study.

Model scoreboard

Independent projects
Churn predictionAccuracy, PySpark Random Forest
Clinical trialsF1 score, XGBoost + NLP
Credit defaultAUC, XGBoost vs. neural network

Measured improvement by role

Percent change, scale 0–40%
  • AptechModel accuracy+20%
  • AptechLatency for 50,000 users, lower−25%
  • SinebyteChatbot latency, lower−30%
  • DeepKoreData quality, automated SQL validation+25%

Credentials

Education & certifications
  • M.S. Business Analytics and AIUT Dallas · Scholar with Recognition3.80 / 4.0
    2023–2025
  • B.Tech Computer ScienceVellore Institute of Technology · Scholar with Distinction3.54 / 4.0
    2019–2023
  • Azure DP-100
  • AWS Cloud Practitioner
  • SAP S/4 HANA
  • SAP Analytics Cloud

About

Models that ship, and numbers that hold up

I'm a data scientist at Aptech, where I build PySpark and SQL models on healthcare software datasets above one million records, and ship the ETL pipelines and REST microservices that put them into production.

Before that I built a GPT-based chatbot on Flask and React at Sinebyte, and worked with Ford Motors on connected-vehicle safety data at DeepKore, writing SQL pipelines and Tableau KPI dashboards. I hold an M.S. in Business Analytics and AI from UT Dallas and a B.Tech in Computer Science from VIT.

My habit across all of it: define the metric first, measure before and after, and report the change in plain terms.

  1. Sep 2025 – present

    Data Scientist, Aptech

    Healthcare software · PySpark, SQL, REST APIs

    +20% accuracy · 99.5% integrity · −25% latency

  2. Jan – May 2025

    Software Data Engineer, Sinebyte

    IT startup · Flask, ReactJS, GPT

    −30% chatbot latency

  3. Sep 2022 – Jul 2023

    Data Science Intern, DeepKore

    AI-based SaaS startup · SQL, Tableau, for Ford Motors

    +25% data quality

Case studies

Business problem, approach, measured outcome

Seven pieces of work, four from jobs and three from independent projects. Pick one to open it.

Work · Aptech, healthcare software · Sep 2025 – present

Accurate models and fast services on 1M+ healthcare records

Problem
Healthcare software teams needed predictions they could trust on datasets above one million records, and services that stay responsive for 50,000 users.
Approach
Built and deployed ML models with PySpark and SQL. Architected ETL pipelines and integrated microservices into production through REST APIs.
Outcome

Accuracy up 20%, 99.5% data integrity, latency down 25%.

Data integrity, 1M+ records99.5%
Latency reduction, 50,000 users25%
Model accuracy gain20%
  • PySpark
  • SQL
  • ETL
  • REST APIs
  • Microservices

Project · Aug – Sep 2026

Real-time ASL transcription for video calls

Problem
Signers on enterprise video calls have no built-in way to be read live without a human interpreter or typing.
Approach
A PyTorch and MediaPipe gesture classifier, with a 20-frame memory cache and an 85% confidence gate so noisy frames don't produce false text. An OpenCV pipeline broadcasts the predicted text through a virtual webcam.
Outcome

Live predicted text inside a video call at 30 FPS.

  • 30FPS low-latency video pipeline
  • 85%confidence gate removes false positives
  • 20frames of memory caching
  • Python
  • PyTorch
  • MediaPipe
  • OpenCV

Project · Oct – Nov 2025

Clinical trial outcome prediction

Problem
Phase 3 trials are expensive, and teams want an early signal of which ones are likely to succeed.
Approach
An XGBoost model on the TDC dataset that combines molecular fingerprints with NLP keyword features from trial text, built in a hybrid feature-engineering pipeline.
Outcome

0.78 F1, and clinical text carried strong feature importance.

F1 score, scale 0 to 10.78
  • XGBoost
  • NLP
  • Feature engineering

Project · Jun – Jul 2025

Customer churn prediction at Spark scale

Problem
Retention teams need to know which customers are likely to leave before they do.
Approach
A PySpark pipeline that preprocesses 7k+ customer records, engineers features and predicts churn with a Random Forest. Cross-validation and AUC checks run in Spark's ML library.
Outcome

84% accuracy, with a faster ETL step.

Accuracy84%
  • PySpark
  • Random Forest
  • Spark ML

Project · Credit risk

American Express credit default prediction

Problem
Lenders need to flag likely defaulters early from long customer statement histories.
Approach
Trained and compared an XGBoost model against a neural network on the default-prediction task.
Outcome

0.94 AUC separating defaulters from non-defaulters.

AUC, scale 0 to 10.94
  • XGBoost
  • Neural networks
  • Risk modeling

Work · Sinebyte · Jan – May 2025

GPT-based chatbot that answers 30% faster

Problem
A chatbot has to respond quickly to feel usable inside a web product.
Approach
A Flask and REST API backend with a JavaScript client, plus a responsive, modular interface built in HTML, CSS and ReactJS.
Outcome

Latency down 30% and a smoother user experience.

Latency reduction30%
  • Flask
  • REST APIs
  • ReactJS
  • GPT

Work · DeepKore, for Ford Motors · Sep 2022 – Jul 2023

Connected-vehicle safety KPIs for Ford

Problem
Ford's large-scale user safety data from connected vehicles needed trustworthy KPI tracking.
Approach
SQL pipelines and Tableau dashboards for KPI tracking, automated SQL validation, and predictive modeling to identify active users.
Outcome

Data quality up 25% through automated validation.

Data quality improvement25%
  • SQL
  • Tableau
  • Data validation
  • Predictive modeling

Stack

Tools, with the result each one produced

Languages

Python and SQL carry the 99.5% integrity and 25% data-quality results. JavaScript shipped the chatbot interface.

  • Python
  • SQL
  • R
  • Bash
  • Java
  • Scala
  • JavaScript
  • HTML/CSS

Machine learning

XGBoost and NLP features reached 0.78 F1 on clinical trials. PyTorch and MediaPipe run live at 30 FPS.

  • XGBoost
  • Random Forest
  • PyTorch
  • TensorFlow
  • Keras
  • NLP
  • Generative AI
  • OpenCV
  • MediaPipe
  • Predictive modeling

Data engineering

PySpark pipelines and REST APIs cut latency 25% for 50,000 users. Tableau dashboards tracked Ford's safety KPIs.

  • PySpark
  • Spark
  • Hadoop
  • Hive
  • Databricks
  • Snowflake
  • dbt
  • ETL
  • Flask / REST APIs
  • Docker
  • CI/CD
  • Tableau
  • Power BI
  • Looker