Chaithanya VemuriDATA & ML ENGINEERING

BERLIN, GERMANY / MSc DATA SCIENCE

Hi, I’m

ChaithanyaVemuri.

DATA & ML ENGINEERING

I build data pipelines and compare ML models against simple baselines. I’ve also used a GKE lab to learn how a service recovers from a bad release.

View case studies Download résumé

Open to junior Data & ML Engineering roles

Chaithanya VemuriBerlin, Germany
INDEPENDENT ENGINEERING & MSc RESEARCHSelected work
411.7M
Historical events processedREES46 data pipeline
8 methods
One chronological comparisonOTTO MSc dissertation
GCP / GKE
Delivery & recovery exercisesFulfillAI platform lab

01 / SELECTED WORK

Project case studies

A pipeline that ran out of disk, a GKE release that couldn’t pull its image, and a weekly baseline that beat a Transformer.

REES46 / DATA ARCHITECTURE01
B
Bronze411,709,736 historical events
RAW → PARQUET
S
Silver410,325,314 canonical events
VALIDATE / DEDUP
G
Monthly GoldPartitioned, restartable outputs
FEATURES / FACTS
Then: retrieval → ranking → local API
01 / REES46 V2DATA ENGINEERING · RECOMMENDATIONS

Making a 411.7M-event pipeline restartable

My first aggregation ran out of temporary storage. I split the work by month and session bucket so I could reuse finished work after a retry.

  • Python
  • DuckDB
  • Parquet
  • FastAPI
411.7Mhistorical events processed
Read the case study Source code
FULFILLAI / DELIVERY ARCHITECTURE02
{ }SourceGitHub
→
✓Build & scanActions · WIF
→
◇ImageExact digest
GIT-BACKED DESIRED STATE↓
GKE AutopilotArgo CD reconciles the Helm workload
API LAB
Health → metrics → alerts → recovery
02 / FulfillAIDATA & ML · CLOUD PLATFORM LAB

Rehearsing a bad release on GKE

I built a pipeline on synthetic fulfillment data, then set up a separate GKE API lab. I released a deployment pointing to a nonexistent image, then verified recovery.

  • PostgreSQL / dbt
  • GCP / GKE
  • Terraform
  • Argo CD
50,000synthetic orders in the data workflow
Read the case study Source code
OTTO / SELECTED CV RESULTS03
MEAN CV WAPE · LOWER IS BETTERWeekly seasonal13.93%Seasonal blend13.99%Extra Trees14.35%GRU15.42%Transformer17.67%0%10%20%
Selected on CV, before the final holdout
03 / OTTO forecastingMSc RESEARCH · TIME-SERIES FORECASTING

Why a weekly baseline won on OTTO

I compared eight forecasting approaches on four weeks of hourly cart and order counts. Repeating last week’s pattern gave the lowest mean WAPE on the development folds.

  • Python
  • PyTorch
  • scikit-learn
  • Time-series CV
8approaches compared chronologically
Read the case study Source code
Explore all case studies

02 / HOW I WORK

How I work

What changed after I ran out of disk, compared forecast errors, and tested a bad release.

Check the data first

In REES46, splitting the work only helped if every event in a session still went to the same bucket.

Pipelines & data quality

Keep a baseline

For OTTO, I chose the model on chronological development folds and kept the last 96 hours for the final test.

Evaluation & reproducibility

Test the rollback

In FulfillAI, I checked how Kubernetes behaved when it couldn’t pull the new image and how to get back to a healthy release.

Delivery & operations labs

03 / BACKGROUND

From electronics to data engineering

I studied Electronics & Communication Engineering before moving into Data Science. I’m now pursuing an MSc at Gisma University of Applied Sciences.

Most of my technical work so far comes from independent projects and my dissertation. The case studies link to the code and the experiment records.

More about my background

OPEN TO JUNIOR OPPORTUNITIES

Get in touch

I’m looking for junior Data and ML Engineering roles. If you have a role in mind or a question about the work, I’d like to hear from you.

Contact me