I build AI agents and generative pipelines in production.

Agentic AI systems in production: RAG, LLM evaluation, and a GCP data platform, with real metrics.

• 70→90%

  validation
• −75%

  data cost
• 3.6M

  records

• Multi-tenant conversational assistant with an agentic architecture (orchestrator and sub-agents) on Google ADK, with multi-provider routing via LiteLLM configurable without redeploy.
• Automated RAG quality evaluation subsystem (LLM-as-judge, multiple dimensions) with dataset generation at ingestion; fixed a defect that cut false negatives from 91% to nearly 0%.
• Document validation pipeline via multi-LLM consensus (GPT + Claude + Gemini) with OCR: validation pass rate from 70% to 90%.
• GCP data platform (BigQuery, Dataform, Dataflow; medallion architecture): 3.6M+ records processed with zero data loss, backfill 2× faster (34 → 17 min) and orphan-document cleanup; process cost reduced ~75% (from ~€1,400 to ~€300).