Xavier
Mareca

Data & AI Consultant

Practical data and AI systems. Without the hype.

View Projects
LinkedIn GitHub X
0
Years in BI
0
Projects Delivered
0
AI Apps Built
0
EU Languages
Background

8 years reading data.
Now teaching machines to do it.

8 years in data — Power BI, DAX, and semantic models across finance, retail, and logistics. When LLMs became viable for production work, I moved into RAG pipelines, LangChain, and multi-agent systems with CrewAI. The projects here sit at the intersection of both: pipelines that feed models, models that enhance dashboards, agents that automate the layer in between.

AI & LLMs
RAG pipelines, LangChain, CrewAI agents, Anthropic Claude integrations.
Cloud Engineering
AWS SageMaker, Lambda, Step Functions — serverless, event-driven architectures.
Business Intelligence
Advanced DAX, semantic models, high-performance Power BI analytics.
Spanish · French · English · German · LinkedIn
Current

Current focus

AWS MLA‑C01 · DP‑700
ML Engineering + Microsoft Fabric — dual exam track, 2026
In Progress
Multi-Agent Systems
CrewAI and LangGraph for complex AI workflows
Researching
Anthropic Claude Integration
Production RAG systems with LangChain and ChromaDB
Completed
Work

Selected projects

AI-Powered BI Assistant

Natural language querying over structured data. RAG pipeline with full test suite — 86 integration tests with pytest.

LangChainRAGClaudeFastAPIChromaDB
v1.0.0
• Handles complex JOINs and aggregations
• CI/CD with GitHub Actions
xdt-ai/bi-assistant

HR RAG Assistant

Multi-turn conversational assistant over HR policy documents. Session memory, automatic re-indexing on document updates.

LangChainChromaDBClaudeFastAPIGitHub Actions
CI/CD
• ConversationalRetrievalChain with memory
• Docker + GitHub Actions pipeline
xdt-ai/hr-rag-assistant

Sales Analytics Pipeline

Multi-source ingestion (ERP, CRM, FX API) through a medallion lakehouse to a Kimball star schema and Power BI dashboards. Dirty records quarantined, not silently dropped.

PySparkDelta LakePower BIAzureDatabricks
Medallion arch.
• Kimball star schema: fact_sales + 3 dims with FX-to-EUR conversion
• pytest + CI smoke-run on every push
xdt-ai/sales-analytics-pipeline

Document Intelligence

AWS Textract handles OCR and layout extraction; Claude structures and validates the output. Configurable extraction schemas per document type, with per-field confidence scores.

ClaudeAWS TextractPython
Structured output
• Supports PDF, PNG, JPG
• Configurable extraction schemas per document type
xdt-ai/document-intelligence

AWS Data Platform

End-to-end lakehouse: raw data lands in S3, PySpark jobs on Glue and EMR transform it into Parquet across zones (raw → staged → curated), and curated marts load into Redshift for BI.

PySparkGlueEMRAthenaRedshiftTerraform
Medallion lakehouse
• S3 raw → staged (Parquet, partitioned by ds) → curated
• Athena for ad-hoc SQL · Redshift as serving layer for BI
• IaC with Terraform, isolated dev/staging/prod environments
xdt-ai/aws-data-platform xdt-ai/aws-sagemaker-hpo-lab4 — SageMaker HPO (XGBoost)
Tools

Tech stack

Business Intelligence
Power BI
DAX
SQL Server
Semantic Models
AI & Agents
CrewAI
LangChain
Anthropic Claude
RAG
Ollama
Vector DBs
Data Engineering
PySpark
Delta Lake
Databricks
SQL
Python
Pandas
Cloud & APIs
Azure
AWS
Microsoft Fabric
FastAPI
Docker
GitHub Actions