What clients found
when they worked with us

Challenge

The facility had experienced four unexpected cooling system failures in eighteen months. Each resulted in product loss. The maintenance logs were detailed but the team had no way to identify the sensor patterns that appeared before each failure — the data existed but was not being read systematically.

What We Did

We worked with four months of temperature, pressure, and vibration sensor data alongside the maintenance logs. The model was trained to identify patterns from the thirty-six hours before each historical failure. Validated against the most recent failure period (held out from training).

Result

The model correctly flagged the held-out failure 28 hours before the recorded incident. False alarm rate during validation: one alert over the validation period that did not precede a failure. The facility has since deployed the model. No unexpected shutdowns in the three months following deployment.

Challenge

A team of eight coordinators produced daily route summary reports and weekly consolidated dispatching memos. Each person had a slightly different writing style and the reports were taking 20–30 minutes per person per day to write — time that could be spent on coordination tasks.

What We Did

We reviewed sixty historical reports across both document types and built a helper trained on the team's combined writing patterns. The tool accepts the key data inputs through a short form and produces a draft. Three review rounds with the team before handover.

Result

Daily route summaries now take five to eight minutes per person including review. Weekly memos still require fifteen to twenty minutes of editing — the team expected this based on the limitations guide. Overall, the coordinator team estimates about ninety minutes per day saved across the group.

Situation

A mid-sized distributor wanted to explore whether an AI forecasting model could reduce overstock. They had heard of demand forecasting tools and assumed their ERP export history would be enough to build one. The initial conversation was booked to assess this.

What We Found

The ERP data covered eleven months — shorter than the seasonal cycle for their product categories. The demand patterns also showed three structural shifts in the last two years due to supplier changes, making historical patterns a weak basis for forecasting. We outlined what would be needed to build something useful.

Outcome

We recommended waiting until they had 18–24 months of clean data from a stable product range before investing in a forecasting pilot. No project followed. The client has since said the assessment saved them from commissioning work that would not have produced a deployable result.