Construction
Mosaic estimates a job in a third of the time.
A model that reads the drawings and quantities, an estimating agent and a draft estimate the estimator corrects.
-38%less time producing an estimate
±6%deviation from final cost
Manufacturing · Sand
Sand's engineers stopped guessing why a line stopped.
A system that watches the production line and explains a stoppage before an engineer has to ask.
What changed
- less time to find why a line stopped
- -70%less time to find why a line stoppedminutes from stoppage to a logged cause, 3 lines, 8-week average
- stoppages logged automatically
- 100%stoppages logged automaticallyshare of stoppages with a system-written entry, no manual typing
At a glance
- Starting point
- A pilot that must reach production
- Capabilities
- Detect
- What we built
- A detection model on the existing sensors, an explanation agent and a stoppage log the engineers read on their desk.
- Where it runs
- On-premise, next to the line controllers. Sensor data never leaves the plant.
The problem
When a line stopped on Sand’s factory floor, an engineer had to leave what they were doing, walk to the machine, check the sensor readings and the error log, and figure out what happened. By the time they arrived, the context was often gone: a jam had cleared itself, a reading had reset, and the cause stayed a guess.
Supervisors wrote the stoppage down by hand at the end of the shift, from memory. The log rarely matched what had actually happened.
What we built
We connected the existing sensors to a system that reads vibration, temperature and speed data in real time, flags a stoppage the moment it happens and matches it against the patterns of past incidents. It writes a short, plain explanation an engineer can act on immediately, with the sensor readings attached.
The system does not decide what to fix. It gives the engineer the same picture they used to spend twenty minutes reconstructing, the moment the line stops.
How it runs in production
We started on one line, running the system alongside the old paper log for a month so the maintenance team could check every explanation against what they found on the floor. Only once the two agreed on every stoppage did we roll it out to the other two lines.
When the model cannot match a stoppage to a known pattern, the entry is marked “unexplained” rather than guessing, and the engineer’s own note on that entry becomes the label for the next training pass.
What changed
Engineers now find the cause of a stoppage 70% faster than before, measured over eight weeks across three lines, and every stoppage is logged automatically, with no one typing it in from memory at the end of the shift.
We can see which line stopped and why, before the shift ends.
The maintenance team now spends its time fixing machines, not reconstructing what happened to them.
What we built
The four layers of this system.
For the technical reader
- Stack
- Python ingestion on the plant network, TimescaleDB, a small web log the maintenance team opens on the floor.
- Models
- A time-series anomaly model per line trained on vibration, temperature and speed; a rules layer maps anomalies to known incident patterns.
- Data
- Existing PLC and sensor readings at one-second resolution; the paper stoppage log of the previous year, transcribed as labels.
- Evaluation
- One month of side-by-side running against the paper log on the pilot line; the roll-out started only when the two agreed on every stoppage.
- Infrastructure
- Runs on one industrial PC per plant next to the line controllers; no cloud dependency.
- Security
- Sensor data never leaves the plant network; the log is read-only for everyone except the maintenance lead.
- What we rejected
- Cameras on the line. The sensors already there explained every stoppage the team cared about, at a fraction of the installation and privacy cost.
Next
The next case.
Read the case
Talk to us about a similar problem.
Tell us how it is done today and what changes if it works. You get a straight answer on whether AI can solve it.
Talk to usor write to hello@maistik.studio