Cathode lines run 24/7. Every stop is expensive, and fixing it depends on a few experienced people.
Cathode material for EV batteries moves through mixers, kilns, saggar-handling robots, conveyors, mills and fillers. These lines are long, tightly coupled, and controlled by tens of thousands of PLC instructions and robot jobs.
No safe place to test
Changing a PLC rung or a robot job means trying it on a live line. Vendor simulators replay hand-made motion, not the program that actually runs.
Troubleshooting is tribal
When a robot freezes while holding a saggar, finding the cause depends on who is on shift. The knowledge lives in people, not in a system.
Data without foresight
Alarm logs, heater and motor trends, and maintenance records already exist. They are rarely connected, so failures are found after they happen.
One pipeline, from control code to a maintenance decision
Three layers work today as reference builds. The fourth, a Claude-powered copilot, is what we are building next.
Line Twin
3D twin driven by control logic and process physics.
Fault Lab
Inject faults into the twin and record the symptoms.
Line Monitor
Alarms, trends and failure forecasts in one place.
Claude copilot
Answers technicians using the twin, the fault base and the code.
Line Twin: physics and control logic, not animation
Equipment in our twins is driven by control logic, and powder follows mass balance and physics instead of keyframes. The frames on the left are one simulated mixing batch.
- A five-floor process building, from raw-material hoppers and the mixer down to the roll mill, pneumatic conveying, bag filter and jet mill
- A 2.5 t batch simulated as 20,000 powder clusters at 30+ fps, with mass balance closing within 0.01 kg
- Built-in soft PLC on a fixed 10 ms scan, plus interpreters that run robot jobs and PLC programs unmodified
- Checked against reality: robot cycle time is 32–33 s in the twin and 31.4–34.9 s in controller logs
Equipment library
Parametric models we built ourselves, ready to drop into a customer's line.
- Distance sensor drops to zero (cable contact)27%
- Distance sensor has no output22%
- Sensor reading is noisy22%
- Sensor reads too close (dust, bracket shift)18%
- Sensor value is frozen11%
Fault Lab: learn the failures before they happen
We break the twin on purpose. A stuck cylinder, a damaged sensor cable, a noisy distance sensor, a separator sheet of the wrong thickness: each fault is injected and we record what an operator would actually see.
- 111 robot fault types across 6 robot tasks, 4,375 simulated trials so far
- A field guide in the order a technician works: what you see → what to check first → how likely
- Traceable: every entry points back to the program step where the fault first shows up
Line Monitor: from logs to early warnings
A monitoring system that brings alarms, process trends and maintenance history into one place, then predicts which equipment is likely to fail in the next 24 hours. Operators confirm or reject each call, and the model learns from them.
- Alarm and downtime analytics with top causes, recorded actions and a daily report generated as an editable PPTX
- Heater and motor trends in 10-second buckets, with rule-based trend alarms
- Guarded models: a forecast model is promoted only after it passes recall, specificity and precision gates
One engineer, shipping at team speed. Next, Claude goes inside the product.
Robot and PLC interpreters, a physics-based 3D twin, an ML pipeline and a full-stack monitoring app were built by one founder working with Claude. The next step is putting Claude in front of maintenance technicians.
Claude Code as the engineering team
Claude Code wrote and refactored most of the codebase: the robot job interpreter, the PLC scanner, the fault-injection harness, and a monitoring app of about 37,000 lines.
Reading legacy industrial code
Claude cross-checks PLC comments against robot jobs to work out what each signal means. That is how we reconstructed and verified the robot ↔ PLC handshake.
Maintenance copilot
A technician describes a symptom in Korean or English. Claude reads the live twin state and the Fault Lab knowledge base and returns a ranked check list with its sources.
Explain-the-program
"Why is the robot waiting here?" answered by tracing the actual PLC logic and robot job, not a generic manual.
Shift reports that write themselves
A day of alarms, trends and forecasts turned into a short handover note with prioritized actions.
- Distance sensor cable at the wrist bend: reading drops to zero when the arm moves (27%)
- Sensor power and analog input: reading stays at zero (22%)
- Shielding and grounding: reading jumps while the arm is still (22%)
Long industrial context
Control programs are long and cross-referenced. Claude can hold whole program sections and their comments at once.
Grounded answers
Claude calls tools for twin state and fault records, so each answer cites where it came from.
Korean on the floor
Technicians ask in Korean, engineers document in English. Claude works in both without a translation step.
From reference build to first customer
We are packaging the reference builds into a demo we can sell, then building twins customized to each plant, starting with the equipment every cathode line has.
Demo version
Line Twin, Fault Lab and Line Monitor packaged as one demo.
Sales for custom builds
Meet cathode and battery-material plants and scope twins for their lines.
First custom project
First paid twin built on a customer's own programs and data.
Claude copilot
Maintenance copilot and program explainer in customers' hands.
Built from inside a cathode plant
Sunghun works on a cathode active material production line, alongside the kilns, robots and conveyors this software models. Every feature in CamTwin comes from a real stop, a real alarm or a real troubleshooting session on the floor.
That domain knowledge is the moat. We know which signals matter, what field data actually looks like, and what a maintenance technician needs at 3 a.m.