Romain Futrzynski

Technical advisor

Keeping 700,000 spare parts moving

Problem

Siloed data and hard-to-onboard pipelines across warehouses managing more than 700,000 spare parts.

Approach

I was technical advisor on the migration to Spark Declarative Pipelines, running cadences with the customer’s engineering team.

Outcome

Pipeline latency cut by 99%; usage among business users doubled after one year.

• −99% pipeline latency
• 2× business users
• 700k+ spare parts

• Spark Declarative Pipelines
• Delta Lake
• Databricks

ML methodology

Reading ultrasound like a dermatologist

Problem

Classify dermatologic ultrasound images of skin lesions as benign or malignant.

Approach

I contributed the ML methodology: data preprocessing, feature engineering, model selection and optimizer tuning.

Outcome

77% accuracy, on par with an experienced dermatologist (74%). Published in Skin Research and Technology.

• 77% model accuracy
• 74% dermatologist

• Computer vision
• Transfer learning
• EfficientNet

Returns forecasting without the spreadsheets

Problem

Warehouse staffing for peak periods was planned from manual spreadsheet guesses.

Approach

I advised on replacing them with a Prophet time-series model, right-sized to the problem.

Outcome

The operations team plans capacity from data instead of intuition.

• Prophet
• Time-series forecasting

Contributing author

Writing the exam on GenAI and ML engineering

Problem

Teams adopting GenAI need a shared bar for production quality, from evaluation to deployment.

Approach

As an AI subject-matter expert, I co-wrote the Generative AI and ML Professional certifications and contributed to the Generative AI Engineering course, including the evaluation lifecycle.

Outcome

Engineering best practices codified into certifications and course material.

• MLflow
• LLM evaluation
• Agents
• Vector Search