A distillery is two businesses with two clocks. The still house runs in hours, with physics that has been understood for a century. The warehouse runs in years, with wood, weather and a nosing glass as the final judge. AI behaves very differently in each, and most of what gets pitched to distillers ignores the difference.
This series takes one problem at a time, shows the data engineering that makes it reliable, and puts GenAI in the one role it is actually good at: understanding the question, calling tested tools and explaining the answer.
The series
- A Physics-Informed Digital Twin of a Pot Still: Why Pure ML Twins Drift: Rayleigh distillation, alcohol conservation, and a hybrid model with physics for structure and data for parameters.
- Soft Sensors and Proactive Bands Instead of Alarms at the Spirit Safe: strength between hydrometer readings, bands built from good runs, and fewer alarms rather than more.
- An LLM Copilot for the Still Operator: RAG Over SOPs, Read-Only Tools, No Setpoints: procedure search with citations, handover drafts, and the safety of the tool that does not exist.
- Cask Inventory as Data Engineering: SCD2 Positions, Regauge Events and Angel’s Share as a Derived Number: twelve years of history kept, and loss calculated rather than stored.
- A GenAI Blending Assistant With the Maths Done in Code: a linear programme for the vatting, density tables for the water, and the age statement as a hard rule.
- Where Distillery AI Breaks: A Twelve-Year Feedback Loop: slow feedback, few comparable casks, the panel as arbiter, and what new tools can and cannot change.
Read them in order. Every post is tagged #still-and-model. The wine companion is The Cellar Ledger, and the full catalogue lives on the Distilling & Maturation track.