Family-owned since 1976, K&L Wines sells wine and spirits from a catalog of hundreds of thousands of records. Every day, its team matched messy wine names to the right SKU by hand, and the work grew with every new bottle. With no machine learning engineers in-house, K&L brought in NineTwoThree.
Narrows the full wine catalog to about 50 likely matches.
Reads vintages, shorthand, and misspellings to pick the best.
Type any version of a wine name, get ranked matches back.
Staff confirm the right SKU from 2 to 5 options.
The data was all there. Finding the right record in it was the problem, and the tools K&L had couldn't solve it.
The same wine could appear as "Château Petit 1989," "Chateau 89," or with "Chateau" misspelled. Ctrl+F and database filters couldn't keep up with that many variations across the full catalog.
ChatGPT or Claude can read a messy wine name well, so an LLM became part of the answer. But no LLM can hold hundreds of thousands of records at once, and forcing it to try makes results slow, costly, and unreliable.
With seven people losing a few hours a day each to lookups, K&L couldn't spare the time to second-guess results. A near-miss means the wrong wine, record, or price, so every match had to be right, with a person still confirming it.
The team combined traditional ML and modern LLMs in a way neither could work alone, splitting the problem into two stages.
Custom ML models process all 500,000 records using vector embeddings, Levenshtein and Jaro-Winkler matching algorithms, and pattern recognition to narrow the field to the most likely matches.
With only 50 records within context limits, an LLM applies human-like reasoning. It understands that "Ch." means "Château," "89" means 1989, and recognizes common misspellings.
Instead of handing over a repository, we built an interface where staff type a wine name and get results. They pick the correct SKU from ranked, ML-powered options. No training was needed, and the team started using it on day one.
1 sec
to match 10 wines
14 hours saved daily at $75/hour across 260 working days eliminated the equivalent cost of two full-time employees, totaling over $273,000 in annual savings.
The hybrid ML and LLM approach delivered 99% accuracy validated in production, eliminating the manual errors that came with human matching across half a million records.
Matching 10 wines went from 1 hour to 1 minute. Staff type a name, get 2 to 5 accurate matches, and select the correct SKU with human verification on ML-powered options.
The solution paid for itself in the first month. The simple interface meant immediate adoption with zero change management, and the system scales automatically as inventory grows.
.svg)
NineTwoThree AI Studio’s work led to increased throughput and fewer manual errors in critical workflows. The team managed the project with regular meetings, responded well to needs, and quickly understood business goals.
Build a new platform, or add AI capabilities to existing software, we deliver solutions that work in real operational environments.
Meet with founders Andrew Amann & Pavel Kirillov
Custom AI implementation roadmap
No sales pressure
