How Fresh Blends Made Its Fleet Answerable
A read-only box on their own infrastructure turned a large, distributed beverage fleet into something leadership and operations can simply ask: which machines are about to fail, which were never cleaned, which have gone quiet.
Fresh Blends runs beverage machines. A lot of them, in a lot of places, operated by a lot of different partners. The front of the machine is genuinely good work: their self-serve kiosk won Best Kiosk Innovation at the 2025 ARKI Awards. I was hired for the part nobody photographs: the data behind the fleet.
If you have ever tried to answer a simple question across a fleet like that, you know the problem is not the machines. It is that the answer lives in a dozen systems, and by the time someone has stitched it into a report, the question has moved on.
I built them a data box to fix that. This is what it lets them do now, told straight.
The problem was never "not enough data"
A fleet this size produces an enormous amount of data. The trouble was never volume. It was reach and latency: the signal sat in operational systems no manager was going to query directly, and the reporting that did exist arrived late and flattened into an average.
So the questions that matter went unasked, because asking them was a project. Which machines are about to break? Which ones have not been cleaned? Which have gone quiet? All answerable in principle. None answerable in a sentence.
What we built
One box. Theirs, on their own infrastructure, single-tenant, never shared with anyone else. It connects read-only to their existing databases, where the data already lives, and exposes it through MCP so leadership can ask in Claude, operations can work in Excel, and engineers can drop to SQL. Every query it answers is logged. It was live in days, not months. The software came to their data, not the other way around, and nothing was copied out to a cloud of mine, because there is no cloud of mine in the path.
From "what happened" to "what is about to happen"
The shift that matters is from reporting to prevention. A monthly report tells you a machine stopped selling last week. The box tells you it is trending toward failure now, while there is still time to send someone.
The telemetry was always there. The machines report on their own hardware constantly. What was missing was a way to ask the fleet a question and get the machines that need attention, today, in order. Now that question is a sentence, and increasingly it is an alert that arrives before anyone has to ask.
Concretely, these are things Fresh Blends could not see on demand before, and can now:
| The question it answers in a sentence | What that prevents |
|---|---|
| Which ice makers are trending toward failure? | A machine going dark before anyone schedules a fix |
| Which units are throwing sensor or temperature faults? | Spoiled product and silent, unexplained downtime |
| Which machines are drifting out of calibration? | Off-spec drinks and quiet quality complaints |
| Which machines are overdue for cleaning, or were never cleaned? | A hygiene lapse becoming a brand and compliance problem |
| Which sites are about to run dry or hold stock that is expiring? | Lost sales and wasted ingredients |
| Which machines have gone quiet? | Dark units that keep costing without earning |
None of these were new facts. They were simply unaskable before, and each one is now a question, an alert, and a work order instead of a surprise.
Cleaning is not an ops detail, it is the brand
One capability deserves its own line, because it is easy to underrate. A beverage machine that has not been cleaned on schedule is not just an operational miss. It is a food-safety exposure and a brand risk sitting in a public place with your name on it. "Are our machines actually being cleaned, everywhere, on time?" used to be a question you could only answer by trusting the process. On the box it is a same-day answer, by site, by operator. Trust, but verify, at fleet scale.
Predictive maintenance, in plain terms
The ice maker is the example I keep coming back to, because it fails in a way you can see coming. It degrades before it dies. The same is true of a sensor that starts drifting or a unit that is slowly falling out of calibration. Each of those is a signal the box now surfaces, ranked, so a service visit is something Fresh Blends schedules on their terms rather than something a dead machine forces on them at the worst possible time. That is the entire difference between reactive and predictive, and it is not magic. It is putting the software next to the data and asking better questions of it.
What changed
Leadership and operations stopped waiting for the monthly report. The uncomfortable and the operational questions both became cheap to ask, so they get asked. In the first month, the box answered more than 1,800 of them.
That is the outcome I am proud of, and notice that it is not a revenue chart. It is that a large, distributed fleet became something a person can interrogate honestly, in their own words, on their own infrastructure, with a record of every question asked. The machines were always talking. Now someone is listening in time to act.
Running a fleet whose data can't leave the building? Tell me what leadership keeps asking that nobody can answer.