Portrait photo of Rahul Rawat

Rahul Rawat

Data Scientist. ML Engineer.

I am a data science and analytics master student at SRH Heidelberg who builds end to end machine learning and data systems. My work spans RAG orchestration, fairness aware credit scoring, and cloud data pipelines on GCP.

About

I build data systems that hold up under scrutiny. From distributed Hadoop clusters to fairness audited credit models, my projects share one habit, honest engineering over impressive demos. I care about anti leakage rules, uncertainty intervals, and limitations sections as much as I care about the headline metric.

Tools I work with

  • Python
  • PySpark
  • SQL
  • GCP
  • BigQuery
  • dbt
  • Airflow
  • LangChain
  • Streamlit
  • PowerBI
  • Tableau
  • scikit learn
  • PyTorch
  • CatBoost

Project 1

Hybrid RAG Orchestrator

A modular RAG system that routes every query to the right execution path before a token is generated.

  • Python
  • LangChain
  • Llama 3.1 8b via Groq
  • ChromaDB
  • HuggingFace MiniLM L6 v2
  • Streamlit
  • Modular RAG on Llama 3.1 8b via Groq for high speed inference and LangChain for orchestration.
  • Custom decision making router classifies user intent into three execution paths.
  • ChromaDB with local persistence for document storage and HuggingFace MiniLM L6 v2 for semantic embeddings.
  • Stateful MemoryAgent maintaining conversational context across turns, integrated directly into the inference pipeline.
  • Shipped behind a Streamlit interface, end to end ownership.

Project 2

CreditIQ Fairness by Design Credit Scoring

Credit scoring that clears the EU AI Act fairness threshold without giving up accuracy.

  • Python
  • scikit learn
  • AIF360
  • SHAP
  • Streamlit
  • Lifted Disparate Impact ratio from a failing 0.79 to a compliant 0.88, clearing the EU AI Act and AGG 80 percent fairness threshold.
  • Diagnosed a hidden intersectional bias where younger women were penalised even after age bias was fixed, using SHAP, then designed a 4 way age by gender threshold matrix to correct it without over correcting.
  • Cut the false negative rate from 44 percent to 16.7 percent while holding accuracy at 75 percent.
  • Streamlit decision support tool with a plain language LLM generated explanation, keeping a human in the loop per GDPR Article 22 and EU AI Act Article 14.
  • Unit tests at 100 percent branch coverage and a full regulatory write up.

Project 3

Real Time Flight Tracking Data Pipeline

Live air traffic from the OpenSky Network enriched, orchestrated, and visualised on Google Cloud.

  • Python
  • PySpark
  • BigQuery
  • dbt
  • Apache Airflow
  • GCP Dataproc
  • GCS
  • Tableau
  • TabPy
  • Real time collection from the OpenSky Network API every 30 seconds, enriching each position against airport, aircraft, and weather data, four sources.
  • PySpark cleaning on Google Cloud, dbt for analysis ready tables, nearest airport computed per aircraft.
  • Apache Airflow orchestration so batch and real time layers refresh every 15 minutes.
  • Tableau plus TabPy findings, air traffic drops 4.4 times in heavy rain and clusters around Frankfurt and Munich hubs.
  • Over 128 thousand records processed.

Project 4

Lake Water Quality Forecasting, eRay GmbH

Recursive time series forecasts for a German lake with honest uncertainty and honest limits.

  • Python
  • CatBoost
  • Prophet
  • scikit learn
  • MICE
  • End to end recursive time series pipeline forecasting chlorophyll a, turbidity, pH, and dissolved oxygen for a German lake.
  • Benchmarked six models head to head, Ridge, Gradient Boosting, LightGBM, XGBoost, CatBoost, Prophet, and used CatBoost multi quantile regression for asymmetric 80 percent prediction intervals.
  • Strict anti leakage rules surfaced the honest finding that pH and dissolved oxygen are physically predictable while chlorophyll a and turbidity are not without live optical sensors.
  • MICE imputation for missing winter readings, synthetic winter decay forecast canvas to stop tree models flatlining.
  • Orchestrator with gate checks that halts on failed imputation rather than letting bad data cascade.

Project 5

Movie Analytics and ML Pipeline on GCP

A fully automated medallion architecture on BigQuery that predicts hits before release.

  • GCP BigQuery
  • Cloud Run
  • GCS
  • Cloud Scheduler
  • BigQuery ML
  • Python
  • SQL
  • Looker Studio
  • End to end batch pipeline pulling movie data from a public API into a GCS data lake, Bronze to Silver to Gold medallion architecture in BigQuery, fully automated on a Cloud Scheduler trigger with no manual steps.
  • Silver layer with schema enforcement, safe type casting, deduplication via window functions, genre normalisation.
  • BigQuery ML classifier predicts whether a film will be a hit before release, features split into two tables to prevent leakage, pre release signals only.
  • Gold aggregates and a five page Looker Studio dashboard on genre ROI, foreign language growth, release season timing, plus an ML early warning view.
  • Least privilege service account and Secret Manager.

Project 6

Hadoop Data Crawling and Processing Platform

A self healing Hadoop cluster on Docker Swarm fed by a decoupled scraping pipeline.

  • Python
  • Selenium
  • BeautifulSoup
  • Pandas
  • Docker Swarm
  • Hadoop
  • HDFS
  • SQL Server
  • Distributed Hadoop cluster with 1 Name Node and 3 Data Nodes on Docker Swarm, automated container management and self healing.
  • Decoupled web scraping pipeline with Python and Selenium for dynamic paginated e commerce results, raw HTML saved locally for data safety.
  • BeautifulSoup parser handled missing data and decoded sponsored click tracking URLs into clean product links.
  • Structured CSVs ingested into HDFS, redundancy tests verified replication across the 3 Data Nodes, processed data extracted into SQL Server for downstream analysis.

Project 7

Fast Food Nutritional Analyzer and Meal Simulator

A Tableau meal simulator where selecting points builds a live nutritional shopping cart.

  • Tableau
  • Set Actions
  • Parameters
  • Calculated Fields
  • Data Storytelling
  • UI and UX
  • Dynamic shopping cart in Tableau via Set Actions, users select scatter plot points and instantly total calories, fat, and protein for a simulated meal.
  • Parameter driven analytics with a dynamic Y axis via a CASE statement tied to a user controlled goal parameter, muscle gain versus weight loss.
  • Complex order of operation IF and THEN calculated fields for logical grouping and custom flags, an Is It A Trap flag for deceptive high fat and high calorie items.
  • Two tier layout, executive macro view and granular food finder, colour blind safe dark mode palette to reduce time to insight.

Project 8

Economic Impact Analysis of Global Climate Events

Predictive analytics turning raw global climate event data into resource allocation insight.

  • Python
  • Pandas
  • scikit learn
  • Matplotlib
  • Seaborn
  • Random Forest
  • Statistical Modelling
  • End to end predictive analytics and BI study transforming raw global event datasets into structured insights for resource allocation and risk assessment.
  • Advanced data preparation and cleansing over outliers, imputation, and normalisation.
  • Random Forest models analysing correlations between disaster duration and financial impact, extracting clear business relevant insights.
  • Communicated complex statistical findings to non technical stakeholders via visual reports.

Project 9

Diabetes Prediction, Bachelor Thesis

Six classifiers compared honestly on a clinical dataset, written up in IEEE style.

  • Python
  • scikit learn
  • Pandas
  • Seaborn
  • Google Colab
  • IEEE Style Paper
  • Full end to end ML pipeline comparing six classifiers on a clinical dataset of 768 patients, 10 fold cross validation and per model confusion matrices.
  • Caught biologically impossible zero values overlooked in the original paper, applied IQR based outlier removal and proper imputation.
  • Chose ROC AUC over accuracy as the headline metric for a 65 to 35 imbalanced dataset to avoid a misleadingly rosy accuracy figure.
  • IEEE style paper with an honest limitations section.

Experience

eRay GmbH

Lake water quality forecasting

  • End to end recursive time series pipeline forecasting chlorophyll a, turbidity, pH, and dissolved oxygen for a German lake.
  • Benchmarked six models head to head, Ridge, Gradient Boosting, LightGBM, XGBoost, CatBoost, Prophet, and used CatBoost multi quantile regression for asymmetric 80 percent prediction intervals.
  • Strict anti leakage rules surfaced the honest finding that pH and dissolved oxygen are physically predictable while chlorophyll a and turbidity are not without live optical sensors.
  • MICE imputation for missing winter readings, synthetic winter decay forecast canvas to stop tree models flatlining.

Certifications and achievements

Certifications

  • NVIDIA, Fundamentals of Deep Learning

    July 2026

    Verify

  • NVIDIA, Building LLM Applications With Prompt Engineering

    November 2025

    Verify

  • AWS Academy Graduate, AWS Academy Cloud Foundations

    July 2025

    Verify

  • SAS Certified Specialist, Visual Business Analytics Using SAS Viya

    May 2025, valid until May 2030

    Verify

  • Google via Coursera, Foundations: Data, Data, Everywhere

    April 2025

    Verify

Achievements

  • Finalist, USAII Global AI Hackathon 2026, Graduate Level

    Advanced to the final round among the top performing teams, recognised for innovation, technical creativity, and applying AI to real world challenges.

    June 2026

Get in touch