Javier García

Contract intelligence platform for aircraft maintenance

IAG (Iberia, Vueling, Aer Lingus) · Capitole Consulting · Jun 2026 – Present

Senior Data Engineer, one of two data engineers in a 10-person squad

The airline group needed to flag overbilling and contract non-compliance in aircraft maintenance (MRO) invoices: 450 contracts checked against millions of invoices and work orders across 3 airlines.

1. Snowflake, APIs, vendor CSV/XLSX, S3 documents
2. Bronze
3. Silver (dbt)
4. Gold (PostgreSQL / pgvector)
5. Knowledge graph and compliance agents

What I built

• Co-designed and built the platform's data foundation from an empty repository to UAT: an S3-to-PostgreSQL/pgvector medallion architecture orchestrated with Airflow on EKS.
• Built the founding Python ingestion pipeline (boto3, SQLAlchemy, Alembic) and made it idempotent on a SHA-256-based document identity: corrections become new versions, retries are no-ops and a re-run never pays twice for the same LLM call.
• Delivered typed Gold models (contracts, price books, scope rules, turnaround commitments, billing lines) linking each invoice to its contract, with Bedrock embeddings and SCD2 history, as the source for the reasoning team's knowledge graph and compliance agents.
• Owned source integrations from Snowflake, external APIs and vendor feeds into airline-scoped Gold tables, and integrated a document-AI extraction gateway into the contract registry FastAPI service.
• Secured the stack with no long-lived credentials (IRSA, short-lived RDS IAM tokens, TLS verify-full) and per-airline row-level security; migrations are proven reversible by a CI round-trip test on real PostgreSQL.

Results

4 releases

from empty repository to UAT in 4 months

6 DAGs

on Airflow, with 57 dbt models and 2,150 tests

83.2% → 91.4%

compliance-rule coverage after a rule review with business SMEs

• Python
• Airflow
• EKS
• dbt
• PostgreSQL
• pgvector
• S3
• Snowflake
• Amazon Bedrock
• FastAPI
• Alembic

Advertising data platform

L'Oréal, European marketing division · Devoteam · Aug 2025 – Jun 2026

Senior Data Engineer, project lead

Campaign performance and social engagement for global product campaigns lived in 7 separate platforms (DV360, Meta, Google Ads, TikTok, Amazon DSP, Snapchat and offline media) and had to be consolidated for reporting across Europe.

1. 7 advertising platforms
2. BigQuery
3. Dataform ELT
4. Star schema
5. LookML semantic layer, Power BI

What I built

• Led the platform consolidating campaign performance and social engagement into BigQuery, with Dataform ELT deployed continuously.
• Refactored the data model into a star schema of 45+ fact and dimension tables, with a rule-based matching engine for quarterly benchmarks.
• Provisioned dev/qa/prod on GCP with Terraform (Cloud Build CI/CD, Workflows, Cloud Scheduler), running daily automated Power BI refreshes.
• Built a LookML semantic layer with 50+ marketing KPIs for self-service analytics.
• Built the offline-media ETL pipeline (TV, sponsorship) and led the migration adding age/gender breakdowns.

Results

−40%

query latency

5 min → 1 min

dashboard load time

50+ KPIs

in the semantic layer, used across Europe

• BigQuery
• Dataform
• Terraform
• Cloud Build
• Workflows
• LookML
• Power BI

GenAI analytics agent for marketers

L'Oréal, European marketing division · Devoteam · Aug 2025 – Jun 2026

Senior Data Engineer, project lead

Marketers needed answers from the advertising data without writing SQL or waiting for a dashboard.

1. Marketing question
2. Agent on internal ChatGPT (OpenAI APIs)
3. RAG data model (Vertex AI)
4. BigQuery query
5. Answer

What I built

• Designed a GenAI analytics agent on an internal ChatGPT deployment, with a RAG data model in Vertex AI that routes marketing questions to BigQuery queries.

Results

500+ users

in production across the division

• OpenAI APIs
• Vertex AI
• RAG
• BigQuery

Cyberdefense data platform for insurers

Insurers including Mapfre · Evolutio Cloud Enabler · Aug 2024 – Jun 2025

Senior Data Engineer

A cyberdefense product for insurers needed high-volume API data ingested in batch and real time, and served quickly to the front-end team.

1. API, batch and real time
2. S3 / Athena / Glue data lake
3. Redshift
4. Front end and forecasting models

What I built

• Co-built the Python/Terraform pipeline ingesting millions of rows per hour from an API, in batch and real time.
• Migrated 2 years of historical data from the S3/Athena/Glue data lake to Redshift for the front-end team (PySpark on Glue, EKS, Grafana).
• Defined the technical roadmap for ML and generative AI applications, and built customer features and time-series models (ARIMA, XGBoost) forecasting hourly call volumes on Redshift and SQL Server.

Results

−50%

processing time after the Redshift migration

−25%

weekly data incidents

• Python
• Terraform
• PySpark
• AWS Glue
• Athena
• Redshift
• EKS
• Grafana

Classical ML vs zero-shot GPT for classifying EU research projects

European Commission research data (CORDIS) · NTT Data · Oct 2023 – May 2024

AI Engineer

The European Commission's 130,000+ CORDIS research projects had to be labelled across 7 scientific domains, and the question was whether zero-shot LLMs could replace trained classifiers. This work became my MSc thesis at the University of Oviedo.

What I built

• Built a multi-label text-classification pipeline over the CORDIS projects.
• Benchmarked five classical ML models against zero-shot GPT-3.5 and GPT-4o on 13,473 hand-labelled projects.

Results

≈10% vs 15%

Hamming loss: classical models beat zero-shot GPT on label error

13,473

hand-labelled projects in the benchmark

• Python
• scikit-learn
• NLP
• GPT-3.5
• GPT-4o