At Festo, I’m a core contributor to a production agentic assistant serving 2,000+ employees. I designed LangChain and LangGraph workflows with tool-calling, persistent state, retries, and failure recovery.
I built the evaluation service, using expert-written Q&A sets, LLM-as-a-judge, and regression suites to guide prompt and retrieval changes. This work cut unsupported answers by about 30%, while production RAG tuning improved answer relevance by about 25% on an internal benchmark.
I built Dockerized Python services on AWS Lambda and set up tracing, dashboards, and token-cost tracking. A DynamoDB prompt cache and model tiering cut median latency by about 35% and token cost by about 20%.
Previously, at Barberán S.A., I built ML pipelines and forecasting models for industrial datasets, improving forecasting accuracy by about 10%. My projects include transformer-based time-series forecasting and an NLP classification and generation pipeline using RoBERTa, Flan-T5, and T5.
