Client Cases

CNC spindles run under high load; veterans judged them by sound, and sudden failures idled delivery for 1-2 days.

Ingest spindle temperature, current and alarm codes; build a health-score dashboard tracking anomalies and maintenance priority.

Unplanned downtime down 38%; shifted from post-failure firefighting to proactive parts & shift planning

Manual inspection standards were inconsistent across shifts; miss rate ~4.8%, complaints centered on minor defects.

Deployed a single-station AI visual QA POC with industrial camera + stable lighting, linked to scanner for lot traceability.

Defect accuracy 98%+; miss rate cut from 4.8% to under 0.3%

High-mix low-volume orders scheduled on Excel + gut feel; frequent insert orders forced constant rework and inaccurate due dates.

Mapped orders, BOM, routings and capacity; built a lightweight scheduling tool with Gantt views and insert-order simulation.

Scheduling cut from half a day to 30 min; insert-order impact assessment in 5 min

Manuals, SOPs and repair records scattered across drives and chats; field engineers had to phone experts on every alarm.

Cleaned manuals, SOPs and work orders; built an O&M assistant answering natural-language queries by model / alarm code / symptom.

60%+ self-service rate for common issues; new-hire training cut from 2 weeks to 1

Only saw total power bill at month-end — no idea which line or shift cost most; air compressors idling wasted heavily.

Connected main meter, die-cast machines, furnaces and air stations; built 3-tier energy dashboard flagging anomalies by shift/device.

Monthly power usage down 7.5%; shifted from lump-sum to per-device, per-shift management

Spot welds on body-in-white were sampled manually; cold/missed welds caused downstream rework costing multiples of inspection.

Industrial camera + edge box at the weld station; real-time defect detection with position tagging, diverting via PLC.

Weld miss rate cut from 3.2% to 0.5%; QA throughput up 4x

Site patrols were manual; no helmets, smoking, zone intrusion hard to watch full-time — incidents were only found afterward.

Reused existing CCTV for edge-based real-time recognition; auto screenshots + Feishu/WeChat alerts to safety officers.

Shifted from post-incident review to second-level alerts; monthly violations down 60%

High-mix low-volume SKUs with long cleaning changeovers; gut-feel scheduling caused needless changeovers; peak-season capacity strained.

Mapped changeover matrix (flavor / packaging / allergen constraints); built an optimizer that merges similar SKUs to cut cleaning.

Daily changeovers down 35%; capacity utilization up 12%