PRODUCT · Vertical-industry AI platform
The platform is built on Juheng OS. It connects field data with your industry data and outputs three things: the state now, the trend ahead, and the action to take. That is the "two faces" in our brand — one looking back at accumulated data, one looking forward at what the trend predicts.
They can be delivered as one system, or entered separately around whichever problem hurts most today. Combined, they become the industry solution described below.
A unified entity ontology as the structural foundation, a world model as the reasoning core, and controlled agents as the execution loop — turning industry data into a world state that can be reasoned about, acted on, and audited.
Inspection of high-altitude and hazardous areas for asset-heavy industries, with automatic flight routes doing the data collection — taking people out of tower climbing, height work and large-scale repetitive sweeps.
Inspection imagery, asset registers and operating data are connected, producing defect recognition results, degradation trend assessments and inspection reports ready to be filed.
Each layer answers one specific question. We aren't chasing a bigger model — we're making the system able to answer the whole chain from state to action.
Ingests imagery, positioning and sensor data, IoT, logs, work orders and business systems. The problem isn't where the data is, but turning it into entities and events.
A unified entity ontology describes equipment, process and space in one language, then compresses multi-source data into a current world state that is queryable, reason-able and explainable. This is the backward-looking face of "two faces".
Answers three questions: what happens if we do nothing, what happens if we take a given action, and would a different choice at the time have changed the outcome. This is the forward-looking face.
Task decomposition, plan generation, tool calling, approval and rollback. Agents are not direct controllers — every action passes through reasoning and validation first. This is "beginning": every outcome flows back as the start of the next cycle.
For asset-heavy industries the most expensive cost is not the equipment — it is that someone has to climb up and take a look.
High-precision positioning turns "who is where, right now" into computable data. Layer behaviour recognition and wearable terminals on top, and you get a closed loop from sensing to response. This solution line has been deployed and delivered on industrial sites in the energy and environmental sector.
The plant is modelled in 3D from laser panorama capture, with people and equipment mapped onto the model in real time. What a manager sees is not a dot on a 2D plan but which side of which machine someone is standing on, and on which floor.
Work zones and hazard zones are drawn directly in three dimensions. Integrated with the permit-to-work system, the permitted area is the geofence — when the permit expires the geofence tightens automatically, preventing "the permit ended but the person is still inside".
Boundary breaches, over-occupancy, proximity to hazards, incorrectly worn helmets and harnesses — each alert carries position, time and identity, goes straight to the control room and can trigger an emergency plan. Video behaviour recognition can reuse your existing surveillance, adding no new camera points.
The value of an alert is not that it fired — it is who receives it, what it carries, and what action it triggers.
| Alert type | Trigger condition | On-site response |
|---|---|---|
| Boundary breach | A person enters an unauthorised work zone or hazard zone | Platform and control room alert together; person, position and time recorded, with replayable history |
| Over-occupancy | The number of people in a zone exceeds the configured limit | Immediate evacuation prompt, preventing crowding in confined or high-risk areas |
| Proximity | A person comes closer to a hazard (tank farm, rotating equipment) than the safe threshold | Tiered warning — from an early caution at distance to an alert on approach, leaving time to intervene |
| PPE compliance | A helmet or safety harness is not worn correctly | Backend signalling and the wearer's phone both receive the alert for immediate correction |
| One-touch SOS | A person presses the SOS button on their positioning tag | The control room receives the alarm and starts the emergency plan, dispatching to the reported position |
What the system really needs to govern is human behaviour, and people do not carry laptops. Terminal form matters as much as the backend.
Output has to enter your management process directly — otherwise even an accurate algorithm is just another slide deck.
Defects are recognised from inspection imagery and attached to a specific asset and position — which tower, which phase, what defect, how severe.
Data from successive flights on the same route is aligned and compared to assess degradation, upgrading "we found this today" into "how much longer will it last".
Filing-ready inspection reports are generated automatically and connected to your work-order and asset systems, so problems enter the existing handling process instead of stopping inside a PDF.
We prioritise scenarios where inspection is expensive and risky, and where data can accumulate continuously — only those support the value of trend assessment and behaviour analysis.
Complex plant layouts and many hazard zones. Watching by eye and assigning blame afterwards never answers "who is inside a hazard zone right now"
Real-time control through 3D positioning and geofences; boundary, over-occupancy and proximity alerts fire at once, work permits drive the geofence, and full position history stays replayable
Lines are long and points scattered; climbing towers is risky; past inspection results use inconsistent criteria and cannot be compared
Automatic routes give full coverage with fixed route and parameters; defects are recognised automatically and attached to tower and phase, producing a comparable historical record
Dense equipment and long inspection lists depend on senior engineers; anomalies are often only found after a trip or a shutdown
Inspection items are structured into asset states and joined with operating data; inspection frequency is driven by state change rather than spread evenly
Large sites and huge component counts make hotspot and blade defects extremely slow to find by eye
Drone-mounted thermal and visible-light capture identifies abnormal modules and locates them to the string; seasonal comparison produces a degradation trend
Coverage is broad and problems are random; the chain from discovery to handling is long
Scheduled automatic routes generate events and connect to the handling process, turning patrol from "did anyone go" into "was it closed out"
Platform and service are not either-or — one settles who holds the capability, the other gets you results today.
Juheng OS is deployed into your environment and run by your team long term. Suited to customers with engineering capacity who want the capability to become their own asset.
We handle the collection, recognition and reporting, delivering results per mission or per year. Suited to customers who need results quickly, or who have no operations capacity yet.
We don't attempt to cover a whole industry at once. The sensible route is to close the loop in one high-value scenario, then replicate sideways.
Not the broadest coverage, but the scenario where the pain is sharpest and data can accumulate most easily.
Settle what an asset is, what counts as one anomaly, and which fields one event contains. This step decides whether anything can be reused later.
Connect positioning, inspection and operating data into a queryable current state, so conditions on site become visible and comparable first.
Then bring in trend reasoning and one controlled agent loop, moving from "we can see it" to "we know in advance, and someone acts".
Cameras answer "can we see it"; positioning answers "do we know where it is". A person appearing in a frame is a different thing from knowing exactly which floor they are on, which side of which machine, and how far they are from a hazard. The two do not conflict — video behaviour recognition can reuse your existing camera points, while positioning supplies the spatial layer that video struggles to get right.
No. The same positioning data also supports attendance, inspection-completion checks, zone dwell-time statistics and visitor movement analysis. That is why we built it into the platform rather than shipping a single-point system — one data set serving several applications is what makes the investment pay back.
Collection and recognition alone delivers a one-off result; a platform delivers a capability that keeps accumulating. The real difference shows up after the first month — historical data from the same route becomes comparable, trend assessment gains a basis, and inspection frequency moves from spreading effort evenly to allocating it by risk.
No. Drones are currently the most efficient way to collect data at height, but the platform is not tied to a collection method — fixed cameras, positioning and wearable terminals, robots and handheld devices all feed into it. The choice depends on your site conditions and how your teams work.
Under a platform licence, data is deployed inside your own environment and does not pass through our servers. Under an inspection service, collected data is handled as agreed once delivery is complete. Where industry data compliance applies, we write the data boundary into the contract.
The safety control layer shows results relatively quickly — once positioning and geofences are deployed and calibrated, alerts can begin. The value of the trend layer needs at least two collection cycles of history. We won't talk our way around that.
Tell us the one problem you most want solved — "who is inside a hazard zone right now", climbing towers to inspect, or the question of when equipment will fail. We'll start there.
