Clear the claims that are sure. Send people only the ones that need them.

A truck leaves the plant. By the store, the count no longer matches the invoice. Weeks later the retailer pays short and calls it a deduction. Now someone has to open it.

At Winland Foods, seven claims arrived on a normal day and 51 on the worst one. Only one in four ever got a verdict, and three in four of those confirmed a real shortage. The queue was asking people to confirm what was usually true.

A model can clear the obvious ones. But a model graded the easy way looks better than it is, and a model nobody in-house can retrain goes stale the first quarter the mix changes.

It runs where your data already lives, built for your team to own after we leave.

• Every night, after the ERP sync lands, a scheduled task rebuilds every signal from the source tables as they stood that night. No signal that leaks the answer survives.
• The model scores each new claim inside the warehouse and writes the scores to one table, once. Every decision can be audited against exactly what the model knew.
• Your automation reads that table in the morning. Claims above the bar clear on their own. The rest reach the team, most likely invalid first.
• The bar lives outside the model, so the business can move it any morning. Set it stricter at month-end close and trade volume for certainty, as Winland did.
• If the night’s data shrinks or a key signal goes missing, the run stops and logs why, instead of scoring on bad inputs.
• A retraining notebook and a drift view, so your data team refreshes the model on fresh verdicts without calling us.

We grade a model the way it will be used. It learns from the past and is tested on the months after.

• An earlier vendor reported a 0.95 for Winland by shuffling every month into one pile, so the model could study next month’s outcomes while predicting last month’s claims. Graded honestly, the score was 0.81. The gap came from the exam, not the data.
• It has to beat the simplest rule, judging each claim by its customer’s track record, which scored 0.70. Retrained at three points in the year, it held 0.83, 0.81 and 0.83.
• Before launch, every historic claim is rescored through the production pipeline and compared with development. Winland’s 7,709 matched with a correlation of 0.998.
• The result: 28% of claims cleared without anyone opening them, 92% of them real shortages on unseen months, and $174K cleared at the first month-end close. Kickoff to first production run took 25 days.

• Manufacturing

  AI for manufacturers, where the count has to be right.

• AI & Machine Learning
• Data & Analytics
• DevOps & Automation