Arjun Menon measured the cost of a stopped machine in people waiting, not only in minutes.
His factory sat outside Pune and produced metal components for several customers. Some machines were new, others had been maintained through more owners than the workshop’s records could easily describe. A conveyor beside the loading area had a fresh weld. A noticeboard listed who was qualified to operate each press.
“Your brochure says predictive maintenance,” Arjun told Yanchuan during their first video call. “I don’t need a machine that predicts yesterday’s fault. I need to know what to check before we lose an afternoon.”
He turned the camera toward a workbench. A technician was sorting rejected parts into shallow bins. The rejects shared a surface mark, but the defect did not appear on every batch.
“Can your model find that?”
“Perhaps,” Mei answered from Shenzhen. “We need to know what images and measurements exist. More importantly, what happens when a warning is given.”
Arjun liked her answer enough to invite an on-site assessment. Chen insisted on a paid scoping visit and a named local engineering contact. Guo’s firm could finance the study, but Arjun would control his production records.
When Yanchuan arrived, the workshop smelled of cutting fluid and warm steel. He had expected a modern assembly line with clean data. Instead he found a mixture of digital records, paper maintenance notes and knowledge carried by workers who could hear a bearing beginning to fail before a sensor crossed its configured threshold.
An older technician named Leela showed him a notebook with dates and marks against machine numbers. She used one mark when a sound had changed, another when a tool had been replaced. The labels did not match the software supplier’s categories.
“Can you translate them?” Yanchuan asked.
She looked at him.
“You can ask me what they mean.”
He apologized. The distinction mattered. She could explain the record and correct it; a remote developer guessing from the marks could not.
“How many hours do you want from me?” Leela asked.
Yanchuan looked toward Arjun, who was watching them. No hours had been allotted for annotating the notebook. Arjun told Leela the scoping visit could use one paid session with her agreement, not her evenings. Only then did she open to the page she wanted them to see.
Arjun said the last automation vendor had proposed replacing several manual inspection steps with cameras. The cameras were expensive, but not impossibly so. The difficult part was the claim that installing them would reduce the need for experienced workers.
“Did the workers agree the inspections were repetitive?” Ning asked.
“No. The vendor said they were.”
Leela took them to a machine whose vibration was within its usual range. She placed a gloved hand on the housing and asked the operator to make a brief permitted run. The noise changed under load. The recorded dashboard stayed green. The maintenance crew had already scheduled an inspection during the next stoppage.
“Don’t tell me to stop listening because your screen likes the number,” Leela said.
The team did not. They prepared an inventory of available sensors and maintenance records, separating observations from diagnoses and diagnoses from completed repairs.
At lunch Arjun showed them production costs. The factory survived on narrow margins. A delayed delivery could make a customer turn elsewhere; a rushed repair could injure a worker or damage an expensive tool. His plan was to improve reliability without turning every decision into a gamble on an opaque system.
“What would success look like?” Yanchuan asked.
“Fewer surprises we could have prepared for. And no one lying to me about what they know.”
That answer was more useful than a percentage.
Back at the hotel Ning drafted a pilot scope that began with one machine group and one fault category. It would help record maintenance observations and identify patterns worth inspecting. The system would not order shutdowns or decide who was qualified to repair equipment.
Guo called and asked whether the proposed service could scale to other customers.
“Maybe after it works here,” Yanchuan answered.
“Will it work?”
“We have not measured the right baseline yet.”
Guo sighed. “Every new customer teaches me what I don’t own.”
On the final day Leela showed them a repaired bearing assembly. She had marked the replacement date in her notebook and entered the corresponding work order in the digital system. The two records agreed because she had linked them, not because the computer had found a way to understand the sound by itself.
The prospect of a useful model seemed smaller. The practical job seemed clearer.
As Yanchuan departed, Arjun handed him a photograph of the whole maintenance team with their permission. Leela stood at the center, refusing to move to the edge for the visitors.
He put the photograph in the project folder, alongside their names and roles. Leela had written the maintenance team’s next shift on the back of the photograph. Yanchuan checked it before proposing the first test date.