Industry LLM for language and orchestration. Specialist models for anomaly detection, remaining useful life, and vision.
IMTS 2026 · McCormick Place · Chicago
Industrial AI that stays on the factory floor.
KUNWU connects machine data, operating knowledge, and governed workflows across equipment, production, and customer service — without replacing the PLC, CNC, or safety systems you already trust.
Sep 14–19, 2026
Chicago
McCormick Place
Governed industrial AI
01 / Company
Shanghai Yixu Intelligence, built for plants that already run.
We build KUNWU, a self-evolving industrial AI platform. It turns scattered equipment events, documents, and shop-floor know-how into a working intelligence layer across three domains: equipment service, production, and customer operations.
KUNWU sits between machines, people, and enterprise systems. Recommendations stay inside expert bounds. Deterministic control still executes. Independent safety interlocks keep final authority.
The name: Kunwu is the legendary ore mountain whose bronze forged blades that cut jade like clay. Data as ore, models as furnace, intelligence forged into working tools.
- 01Close to the processStart from one named machine and one measurable loss.
- 02Connect, do not replaceOPC UA, MQTT, Modbus, PLC/CNC, MES, ERP, CRM and the systems already in place.
- 03Governed actionShadow mode, approval, audit, versioning, and rollback before write-back.
- 04Edge-readyLocal inference and on-site data options when the factory cannot leave the floor.
02 / Platform
From machine signals to governed action.
Not a general-purpose chatbot. An agent platform that senses the floor, understands the business, calls the systems, and learns from the result.
- Sensemachines, events, context
- Reasondata, rules, knowledge
- Acttools, workflows, systems
- Feedbackapproval, limits, outcome
- Evolvecases, models, standards
Equipment, scheduling, kitting, inventory, and service agents that understand, reason, and coordinate with each other.
Knowledge base and retrieval, workflow engine, optimization solvers, API calls, permissions and audit.
Sensors, PLC and CNC, MES, ERP, CRM, and after-sales service systems through standard interfaces.
-
Build the fact base
One continuously updated record of manufacturing reality: equipment state, process parameters, quality inspection, material movement, energy, orders, and cost.
-
Deploy the agents
Predictive maintenance, quality root cause, dynamic scheduling, energy and emissions, and supply resilience agents read live signals and call optimization, forecasting, and SPC tools to reason about them.
-
Close the loop with systems
Bind to MES, SCADA, PLC, AGV, and warehousing so sense, judge, execute, and feed back becomes one governed circuit rather than four disconnected meetings.
-
Review and evolve
Execute, review, learn, evolve. Manufacturing experience is captured, reused, and improved instead of walking out the door with the operator who held it.
A shared language for the plant
Equipment, work orders, recipes, quality events, maintenance cases, and people sit in one governed operational context. Analytics, knowledge, and workflows read the same shop-floor facts.
Equipment
Work order
Recipe
Quality
Maintenance
People
Governed context
03 / What changes
Six problems every plant recognizes.
Faults surface too late
Failures are found after the line stops, so downtime is unplanned.
Fix-when-broken becomes condition monitoring with graded, routed early warning.
Process know-how will not scale
Parameters depend on a few senior operators, and knowledge leaves when they do.
Personal experience becomes a retrievable knowledge base and standard work.
Quality root cause is hard to trace
Defects repeat because causes are found weeks later, if at all.
Post-mortem investigation becomes real-time localization to process and owner.
Scheduling cannot respond
Plans go stale the moment an order shifts, a machine stops, or material is short.
Experience-based planning becomes constrained optimization with human approval.
Energy is managed roughly
Consumption and emissions lack the granularity that cost and reporting now demand.
Flat monthly totals become per-line, per-shift measurement and targeted reduction.
Service knowledge is scattered
Repair history, manuals, and customer context live in separate systems and inboxes.
Case-by-case firefighting becomes an equipment record that compounds into product insight.
04 / Detailed cases
Three closed loops we can walk through at the show.
Eleven high-frequency scenarios run as working prototypes across equipment service, production, and customer operations. These are the three we bring to Chicago.
Equipment service: from alarm to verified repair
The equipment agent is the entry point. Detection, explanation, dispatch, bounded repair, and learning become one record instead of five disconnected steps.
-
Detect
Time-series anomaly detection, visual inspection, and health prediction on top of existing alarms.
-
Explain
Retrieval over manuals, SOPs, and past cases returns a probable cause with the evidence behind it.
-
Dispatch
A digital work order carries owner, priority, spare parts, and progress to the right person on the right shift.
-
Repair
Bounded actions only: reset, parameter restore, process de-rating, bypass — inside an authorized envelope.
-
Learn
Outcome writes back to the equipment history, the case library, and the model that proposed the fix.
Production: when the floor changes, the plan is recomputed
An order moves, a machine stops, material is short. The system does not just flag the risk — it recalculates the plan and shows the reasoning before anything reaches MES.
-
Schedule
Orders, due dates, routings, machine and labor constraints solved together as a multi-objective problem.
-
Kit
Material completeness verified before release; a shortage triggers purchase, transfer, or a schedule change.
-
Predict
BOM, stock, in-transit and lead times resolve into a critical path, a delivery window, and a risk level.
-
Stock
Dynamic safety thresholds, dead-stock warnings, and replenishment proposals tied to the live plan.
-
Execute
A confirmation node stays in front of MES and ERP. Humans approve, then results feed the next run.
Service and growth: knowledge that compounds after the sale
For machine builders, the installed base is the asset. Repair intake, equipment history, and maintenance reminders turn service into a renewal and product-feedback loop.
-
Intake
Repair requests arrive from any channel and land against the machine record, not a generic ticket.
-
Resolve
Product knowledge retrieval supports part lookup, configuration, and field troubleshooting.
-
Maintain
Service history drives maintenance reminders, spare part planning, and predictive contracts.
-
Renew
Usage and service signals surface renewal, upgrade, and expansion conversations before they lapse.
-
Feed back
Recurring field failures return to engineering as evidence for the next product revision.
Running screens
Screens from live deployments. Customer identifiers are masked and interface text is Chinese; English localization ships with each overseas engagement.
05 / Delivery control
See the delivery risk before the customer does.
Order-driven manufacturers lose delivery dates in the gap between ERP, the shop floor, and the customer's chat thread. This platform reads what your systems already know, finds the risk, and sends it to a named person. It does not replace ERP or MES.
What it reads
- ERP orderscustomers, quantities, promised dates
- Schedule detailmodel, surface finish, quantity, routing
- Production actualsoutput, shortfall, stock, shipments
- Tooling and equipmentdie life, trial runs, faults, maintenance
- Conversationscustomer threads, work groups, knowledge base
Three-level delivery risk warning
-
L1
Whole order not scheduled
Near-due orders that have never entered the plan at all.
-
L2
Line item not scheduled
Which model, which surface finish, how many dies are available.
-
L3
Production blocked
Shortfall against plan, material stalled, parts repeatedly re-run.
Customer progress collaboration
- A per-customer production progress sheet pushed on a schedule
- Round-the-clock lookup by keyword, so nobody waits for an answer
- One salesperson can hold many customer threads without falling behind
Key resource intelligence
- Die life expiry with spare-die purchase reminders ahead of the run
- Repeated trial runs, abnormal output, and high scrap surfaced early
- Machine avoidance based on historical scrap and return records
Management briefing
- Daily production report: output, achievement rate, shift performance
- Must-do list, customer radar, and the facts affecting the business
- Chat threads mined into opportunities, to-dos, complaints, and risk
Running screens
06 / Customer operations
Find, qualify, answer, and hold onto the account.
Export sales runs on three failures: prospects chosen by instinct, inquiries answered inconsistently, and accounts that break at handover. The same knowledge base that serves the shop floor can serve the sales desk.
-
Find
Product knowledge becomes search terms, then a list of target companies and named contacts.
-
Research
Public sources build a company profile, purchasing signals, and a read on fit.
-
Answer
The original inquiry plus the knowledge base drafts a specific reply; a person confirms before it sends.
-
Hold
Every exchange lands in the account file as a summary, an opportunity stage, and a suggested next step.
Why it matters for a machine builder
- New hires answer at the standard of your best application engineer
- Specification questions resolve against real product data, not memory
- An account survives the salesperson who opened it
Document compliance review
- Scheduled collection and batch intake of the documents under review
- OCR across mixed formats, with key fields extracted as structured data
- Four-level configurable rules checked by a rule engine and a model together
- Errors located and annotated, so a person reviews exceptions instead of everything
Running screen
The inquiry-reply assistant and document review screens are shown live at the booth rather than published here, because the source captures contain third-party contact details.
07 / Operating discipline
AI needs an operating system underneath it.
The reference plant below ran a four-year World Class Manufacturing program to JIPM audit standard. We include it because it shows what the data layer has to support: a named baseline, losses broken into parts, standard work that prevents the problem returning, and benefits counted to the yuan.
Six pillars on two foundations
Foundation 1 · 5S Clean, sort, set in place, make it visual — so an abnormality is obvious at a glance.
Foundation 2 · Daily management Shift boards and 42 improvement teams, so improvement happens every day rather than every quarter.
Five meeting levels
- MonthlyManagement committee approves pillar progress and resources
- MonthlySix pillar meetings review indicators and cross-department blockers
- Weekly42 improvement teams report progress with on-site coaching
- DailyShift stand-up tracks safety, quality, delivery, and cost
- Rolling5S inspection with red-tag campaigns and closed-loop correction
Three improvements, and how they were found
- Loss
- Switching supply to the fillers meant pushing residual milk out of the line with water. Loss breakdown ranked this the single largest source of milk loss.
- Cause
- Five whys found operators judged push time by feel, and unstable water pressure meant the line was either under-cleared or over-pushed. Pressure stability was the real variable.
- Fix
- A pressure stabilizer, push time set by measured flow instead of experience, the method written into a single-point lesson, and the change rolled across every pre-treatment line.
- Loss
- Fill weight ran deliberately high to guarantee net-content compliance. A few grams per pack across millions of packs a day is a large invisible loss.
- Cause
- Box plots and distribution analysis quantified the variation. A gauge R&R confirmed the measurement was trustworthy, then eight equipment-side sources were isolated.
- Fix
- Clear the pipework, stabilize pressure, lubricate the volume cam, correct seal overlap, calibrate. Target locked at 212.5 g with net-content compliance held at 100%.
- Loss
- Cleaning, sterilizing, heating, waiting for material, and ramping accounted for 18% of all OEE loss — the largest single item of the sixteen.
- Cause
- Video plus roll-paper analysis broke the startup down second by second, exposing people waiting on machines, machines waiting on people, and wasted walking.
- Fix
- Four rounds of eliminate, combine, rearrange, simplify. Internal work moved external, serial steps ran in parallel. Startup loss fell from 18.0% to 11.9% of the total.
Maintenance strategy by consequence
Equipment was graded on safety and quality impact, then given the cheapest strategy that fit. Attention concentrated on the 16% of machines that could actually hurt.
- A · 16%Major safety or quality impact. Condition monitoring plus corrective redesign.
- B · 12%Moderate impact. Time-based maintenance on a fixed cycle.
- C · 72%Minor impact. Run to failure, with operator autonomous maintenance.
- 125 machinesunder condition monitoring across seven methods, including motor temperature indicators, tensioner limit visualization, air-knife current, and filter differential pressure
- 2,000 → 4,000 hblade replacement interval, extended after distribution analysis proved the real service life
- RMB 1.79Mfreed by returning dead spare stock to suppliers, with a further RMB 350,000 cut from opening inventory
What was left behind
On site
08 / Reference outcome
What disciplined operations can deliver.
A four-year World Class Manufacturing program at an anonymized process plant — 440 staff, 22 aseptic filling lines, roughly 690 tons per day. It is the operating discipline our platform is designed to support: baseline definition, loss visibility, standard work, and verified results.
| Indicator | Before | After | Change |
|---|---|---|---|
| Overall equipment effectiveness | 57.0% | 80.3% | +23.3 pts |
| Manufacturing cost per ton | RMB 9,373 | RMB 7,437 | −20.7% |
| Monthly breakdowns | ~300 | 79 | −74% |
| Breakdown time share | 5.5% | 1.01% | −82% |
| Repair cost per ton | RMB 99.92 | RMB 79.50 | −20% |
| Product defect rate | 0.82% | 0.25% | −70% |
| Complaints per 100,000 packs | 0.35 | 0.11 | −69% |
| Material loss rate | 10.6% | 7.45% | −30% |
| Packaging yield | 98.01% | 98.83% | +0.82 pts |
| Annual safety incidents | 36 | 0 | eliminated |
| Water consumption | 6.7 t/t | 5.33 t/t | −20% |
| Electricity consumption | 157.2 kWh/t | 151.4 kWh/t | −4% |
| Natural gas consumption | 38 m³/t | 29.5 m³/t | −22% |
| Labor productivity | 0.37 t/person/day | 0.47 t/person/day | +27% |
- RMB 15.6Mcumulative benefit from focused improvement, against one-off investments measured in thousands
- Beyond 1:100return on improvement spend across the ranked project list
- Zerosafety incidents from year three onward, sustained
Attribution boundary: this is a reference manufacturing transformation case drawn from an anonymized client review. It is not a KUNWU deployment and the figures are not presented as platform results.
09 / How we start
Engineering close to the real workflow.
Customer-embedded delivery begins with one measurable workflow. We connect real data, run a controlled pilot, verify the outcome, and keep only the patterns that can become a reusable capability.
- On-site discovery
- Baseline
- Controlled pilot
- Verified outcome
- Reusable capability
AI recommends and optimizes. Deterministic PLC/CNC control executes. Independent safety interlocks retain final authority.
10 / On the show floor
Meet Starry Wang in Chicago.
Starry founded KUNWU and builds industrial AI systems that connect machine protocols and edge computing with domain knowledge, agents, and enterprise workflows. Mechanical engineering and computer science at Huazhong University of Science and Technology, cloud infrastructure at Intel China, iOS core modules at Apple in Silicon Valley, and platform work inside major manufacturers.
- Industrial connectivity: OPC UA, MQTT, Modbus, REST, events
- Edge and private AI, on-premises deployment, knowledge governance
- Agents for service, maintenance, planning, and shop-floor coordination
Research collaboration. Starry serves as Chief Industrial AI Expert at the Industrial Technology Innovation Center, HUST–Wuxi Research Institute (Huazhong University of Science and Technology). Joint research and industry resources accelerate technical validation and deployment.
11 / IMTS meeting
Reserve 20 minutes in Chicago.
Tell us who you are and what you want to discuss. We reply from hello@yixuai.cn with a confirmed time, usually the same day.