Skip to content

Data Engineer · Xebia

Building the data foundations that analytics and AI run on.

Data engineer at Xebia. I build and run the Google Cloud pipelines behind a global apparel retail brand’s sales and demand reporting, and I work directly with the product owners and BI analysts who depend on it.

Lately that has meant getting the marketing team’s dashboard complete again after a field went missing, making stuck jobs fail and alert instead of reporting success, and moving 524 production schedules onto one shared Dataproc component. I also build RAG and LangGraph agents, at work and in imaarat.ai.

Away from pipelines I play the , rep Real Madrid, and grind way too much Minecraft.

Resume
Live
Just shipped DQ enforcement in the shared pipeline framework: data-quality and profile scans now carry each table’s domain labels
Certified
Xebia Forward Deployed Engineer, FDE-1 Level 2 on GCP, October 2026

01Experience

Xebia

Full-time · 2 yrs 3 mos

Gurugram, India

  1. Junior Consultant, Data Engineer · Client engagement · Global apparel retail brand

    Jul 2026 – Present · 4 mos

    • Data correctness: The site marketing team’s twice-weekly dashboard lost its campaign field in September 2026. I fixed the pipeline and ran a chained dbt backfill over 6 months of events, repairing 1,455 records, so their reports are complete again.
    • Silent failures: Batches stopped at the 4-hour Dataproc limit were still reported as successful. I added time limits to 600+ DAGs, sized from each job’s worst recent runs, so a timed-out job now fails and raises an alert.
    • Pipeline framework: Owned per-task time limits in the shared Vertex AI pipeline framework: each limit comes from measured run times, and the alert names the task that failed. Live on 59 tasks across 25 e-commerce pipelines.
    • Cost: A high-cost Hybris Spark pipeline was cancelled at the 4-hour limit on 9 of 9 days. After I pruned its lookup joins and partition scans, 11 of 16 runs succeeded and daily cost fell 19.7%. I also right-sized clusters for 29 pipelines, cutting their cost 17.6%.
    • Shared component: Merged duplicated Dataproc pipeline templates into one framework component. 524 production schedules now run on it with no change in output.
    • Source migration: Led preprod validation of 7 Hybris pipelines on the new SQL Server Hyperscale source; all 7 matched the old output.
    • Team practice: Reviewed 102 pull requests across 7 repositories, triaged production incidents and wrote 7 design and analysis docs.
    • Dataproc
    • Vertex AI
    • BigQuery
    • DBT
    • Dataplex
  2. Technical Trainee · Centre of Excellence department · Client-focused project

    Aug 2024 – Jun 2026 · 1 yr 11 mos

    • Cloud migration: Led the move of 13 e-commerce sources (OMS, Hybris, Shopify, Salsify, Cordial, Zendesk and more) from Databricks to Dataproc and BigQuery, checking each one against the old output.
    • Order reporting: Designed change data capture for the OMS order tables with Datastream, so order reporting reads one source of truth in BigQuery (snapshot and change views), to replace Qlik.
    • Data quality: Built Dataplex quality scans for 2 data products, with thresholds tested against years of production data, and a CI gate that blocks dbt models shipped without a scan.
    • Client work: Turned the Looker Studio sales and demand reporting needs of the brand’s product owners and BI analysts into data models.
    • Generative AI: At the Centre of Excellence, built RAG document extraction (LlamaIndex chunking into Weaviate and ChromaDB) and a Terraform linter that uses Claude on AWS Bedrock to check all 6 Well-Architected pillars.
    • Apache Spark
    • Dataproc
    • BigQuery
    • Datastream
    • Dataplex
    • LLMs
    • RAG

02How I build

Six rules I work by, each with the work it came from.

records restored to the marketing team’s reports
1,455
records restored to the marketing team’s reports
DAGs that now fail and alert instead of hiding a timeout
600+
DAGs that now fail and alert instead of hiding a timeout
production schedules running on one shared component
524
production schedules running on one shared component
e-commerce sources moved off Databricks, each validated
13
e-commerce sources moved off Databricks, each validated

Reliability

Failures are loud

Our stuck jobs used to hit the 4-hour limit and still report success. Now time limits on 600+ DAGs make them fail with an alert, and per-task timeouts in the shared framework name the task that failed.

Cost

Cost is a design input

Pruning joins and scans on a costly Spark pipeline took it from 0 of 9 runs finishing to 11 of 16, at 19.7% lower daily cost. Sizing clusters for 29 pipelines from measured runs cut their cost 17.6%.

Data quality

Profile first, then enforce

Dataplex thresholds come from years of production data, and CI blocks any dbt model that ships without a scan.

Architecture

Move changes, not tables

Order data moves as changes: CDC with Datastream into BigQuery snapshot and change views, designed to replace third-party replication.

AI

Rules decide, AI explains

In imaarat.ai a deterministic rule engine makes every decision; LangGraph agents read forms and explain the result, and a reviewer signs off.

Operations

Fix the cause

Cleared platform health-check failures at the source: renamed data mart tables to the naming standard and added table and dependent-table descriptions for the data products.

03Projects

imaarat.ai

Underwriters get a risk decision they can check: rules score each commercial property proposal against official seismic, flood and cyclone data for 19,312 Indian PIN codes, AI reads hand-filled forms in 26 languages and explains the score, and a reviewer signs off. I built it myself: hazard joins by PIN code, a nightly ELT to Parquet, 9 dbt models with 26 tests, and LangGraph agents served over MCP and A2A.

  • FastAPI
  • DBT
  • DuckDB
  • PostgreSQL
  • LangGraph
  • Langfuse
  • RAG
  • MCP
Live →Code →Docs →Inspired by a Xebia capstone team project

NYC Taxi Lakehouse

Medallion lakehouse on Delta Lake: bronze ingests 9.5M trips, silver quarantines 3.9% bad records through 5 reject rules, a 7-check quality gate blocks bad loads, and gold dbt marts carry 63 tests.

  • PySpark
  • Delta Lake
  • DBT
  • DuckDB
  • Airflow

IaC Linter

Reviews Terraform against all 6 AWS Well-Architected pillars, security included, using Claude on AWS Bedrock, then applies the fixes and keeps backups.

  • Python
  • AWS Bedrock
  • Terraform

Bookmark’d

Book discovery platform built by a four-member team as our final-year project; I built search, reviews and JWT login in React over PostgreSQL and Redis.

  • React
  • PostgreSQL
  • Redis

04Toolkit

  • Python
  • Apache Spark
  • BigQuery
  • Google Cloud
  • Dataproc
  • Vertex AI
  • DBT
  • Airflow
  • Delta Lake
  • Databricks
  • DuckDB
  • PostgreSQL
  • MongoDB
  • Parquet
  • Looker Studio
  • Gemini
  • Claude
  • LangGraph
  • Langfuse
  • MCP
  • A2A
  • RAG
  • FastAPI
  • Docker
  • Terraform
  • GitHub Actions
  • Git
  • Vercel
  • Jira
  • Confluence

Off the clock

Certificate

Grade 8 in Electronic Keyboard, Trinity Guildhall (Trinity College London)

Earned while learning piano as a child. Open the stage to play an arranger keyboard.

05Let’s connect

Open to data engineering and data platform roles: on-site, hybrid or remote. Email reaches me fastest.