# Physical AI

> Physical AI from Boltcliq: edge perception, robotics and automation, and edge-to-cloud pipelines for farms, mines, ports, plants and utilities across Africa.

Canonical URL: https://www.boltcliq.com/solutions/physical-ai
Part of: https://www.boltcliq.com/llms.txt

Intelligence that sees, decides and acts in the physical world, including where connectivity is poor.
- Farms, mines, ports and plants produce the signals that matter most. Physical AI turns those signals into decisions at the edge, even when the network is gone.
- Edge first: Models run next to the asset, not only in the cloud
- Offline capable: Systems work through outages and sync later
- Safety led: Human override and fail-safe states are designed in
- Embedded: Engineers on site with your operators
### What we do
- Physical AI connects sensors, machines and models into systems that can perceive and respond. A camera over a conveyor belt, a vibration sensor on a pump, a drone over a field, a scanner at a port gate: each one produces data that only has value if it reaches a decision in time. We design the whole path from sensor to action.
- Across much of Africa the path has to work with constraints. Power can be unstable. Mobile data is expensive or patchy. Sites are remote and spares take days to arrive. So we run perception models on devices at the edge, keep a local buffer of events and sync to the cloud when a link is available. A system that needs a constant connection is not a system for a mine at the end of a gravel road.
- We also take the physical world seriously. Anything that moves a machine or opens a valve is wrapped in safety rules, limits and human override. We start with perception and recommendation, prove accuracy on real site data, and only then move toward automated action. Our engineers work on site, with your maintenance and operations staff, until the system is theirs.
### Engineers on the plant floor
- Physical AI projects fail when the team never leaves the office. Sensors are installed badly, labels are wrong and nobody checks the model against what the operator sees. Our forward deployed engineers spend time at the asset. They watch the work, mount the hardware and label data with the people who know what normal looks like.
- That is also how we find the practical limits: dust on a lens, a power cut during a shift, a pump that is serviced on Fridays. Those details decide whether a system works in month six.
- On-site installation and calibration
- Labelling with operators and technicians
- Field tests before any automation
### Capabilities
**Perception and sensing**
- We build vision and signal models that detect defects, count, classify, read gauges and flag unusual vibration or temperature. Models are trained on images and signals from your site, not on generic datasets. Each model reports its confidence so low-quality cases go to a person.
- Vision for inspection, counting and grading
- Vibration, temperature and acoustic signals
- Confidence scores and human review
- Models trained on your site work on your site.
**Edge deployment**
- We package models to run on rugged edge devices with limited power and memory. Updates are signed, staged and reversible. Devices keep working during network loss and report health when they reconnect.
- Optimised models for edge hardware
- Signed, staged updates with rollback
- Local logging and health checks
- A model that cannot run on site is a demo.
**Intermittent connectivity**
- We design store-and-forward pipelines with priorities. Alerts travel first over whatever link exists, including SMS or satellite where justified, and bulk data waits for a good connection. Conflicts are resolved by rules agreed with your team.
- Store-and-forward queues
- Priority channels for alerts
- Conflict-safe synchronisation
- Remote sites are where value is often highest.
**Robotics and automation**
- We integrate with PLCs, SCADA systems, drones, conveyors and handling equipment. We start by recommending actions to an operator, then automate specific, low-risk steps once accuracy is proven. Every automated action has a limit and a manual stop.
- Integration with industrial controls
- Staged move from advice to action
- Manual override and safe states
- Safety is earned in steps.
**Edge to cloud pipelines**
- We build the pipelines that carry events, images and summaries from the edge to storage and analytics. Data is compressed, labelled with site and asset identifiers, and linked to the operational model so engineers can ask questions across sites.
- Asset-linked event streams
- Bandwidth-aware summarisation
- Retention and cost controls
- One site learned, many sites improved.
**Field operations and maintenance**
- We build mobile tools for technicians that work offline, show the asset history and guide inspections. Findings return to the model and improve the next prediction. Maintenance schedules shift from fixed intervals to condition.
- Offline mobile inspection tools
- Condition-based maintenance plans
- Feedback from field to model
- The technician is the best sensor you have.
### Where it applies
- Agriculture
- Large farms and outgrower schemes cannot walk every field. Pests, water stress and harvest timing are noticed late.
- Drone and satellite imagery feed models at the edge and in the cloud. Alerts reach farm managers by SMS even where data is poor.
- The agronomist decides what to spray, irrigate or harvest and when.
- Mining
- Haul trucks, crushers and pumps run in remote sites. Unplanned downtime is costly and spares are slow to arrive.
- Sensors and edge models detect early signs of failure and recommend maintenance windows. Local buffers keep data during outages.
- The maintenance planner schedules work. Safety officers approve any change in operating limits.
- Ports and logistics
- Gates, yards and quay cranes handle containers with mixed paperwork. Queues build when checks are manual.
- Cameras read container and vehicle identifiers, match them to bookings and flag mismatches for inspection.
- The gate supervisor and customs officer decide on every flagged case.
- Energy and utilities
- Substations, solar farms and water networks are spread across large areas with limited staff.
- Edge models monitor equipment and detect faults, leaks or theft. Drones and technicians are dispatched with the right information.
- The control room decides on switching and dispatch. Automated action is limited to safe, pre-approved steps.
### Built on
**Argus**
- Hosts the pipelines and agent workloads that sit behind the edge devices, with portable deployment close to your site.
- Boltcliq product
- Heavy processing can run in-country.
**Zeus**
- Models assets, sensors, events and work orders so readings are tied to the things operators care about.
- Boltcliq product
- A signal becomes a statement about a specific asset.
**Cyclopes**
- Observes devices, models and decisions, and shows drift, outages and overrides.
- Boltcliq product
- You see problems before the shift supervisor does.
### What we do and do not do
- Design and deploy edge perception and pipelines
- Integrate with your industrial and mobile systems
- Run shadow trials before alerts go live
- Train your technicians and engineers
- Document safety limits and overrides
- Replace safety systems or certified controls
- Automate actions before accuracy is proven
- Manufacture hardware or claim certifications
- Require a permanent high-bandwidth link
- Operate your plant for you
### Questions we hear
- What hardware do you use?
- We work with commodity and rugged edge devices, cameras and sensors that your site can support and service. The choice depends on power, climate, budget and local availability of spares.
- Does it work without internet?
- Yes for the core functions. Models run locally, events are stored and critical alerts use whatever channel is available. Data syncs when a link returns.
- How accurate will it be?
- We cannot promise a number before we see your data. We measure accuracy on labelled site data in shadow mode and agree thresholds with you before going live.
- Can the system control machines?
- Only after accuracy is proven and safety limits are agreed. Early phases recommend actions to an operator. Any automated action has limits and a manual stop.
- Who owns the data from our sensors?
- You do. It stays where you choose, including on site or in an in-country facility. Confirm the legal position on data location with counsel.