Fine tuning and agentic pipelines
Agentic automation and a fine-tuned vision-language model for diagnosing print-quality errors, built during two internships at Xerox-Lexmark.
Across two internships at Xerox-Lexmark, I worked on the data and modeling layers behind print-quality diagnosis.
As a Data Engineer Intern, I helped build a pipeline to cleanse raw data into a customer-preferred format using Databricks’ Medallion architecture, Python, and SQL, including migrating data from the Big Decisions platform into Databricks.
As a Data Scientist Intern, I returned to extend that pipeline with agentic automation: a Databricks prototype (PySpark, SQL, Python, LangChain) that streamlined data cleaning across the Medallion architecture, and a fine-tuned version of Microsoft’s Florence vision-language model on Azure ML to classify and diagnose print-quality errors. I tracked experiments with MLflow and tuned hyperparameters with Optuna, and separately researched agentic automation using Claude and MCP servers to design prompt-driven pipelines for data cleaning and ingestion.