# Physical AI system design

Canonical: https://codeatoms.ai/physical-ai-system-design/
License: CC BY 4.0
Publisher: CodeNinja Atoms (https://codeatoms.ai)

Physical AI system design is the engineering of a system that senses a physical operation, reasons over it as typed objects and acts through a named person. Its decisions are what to sense, where inference runs (the edge, the site or one in-country node), the object model, the open-weight model choice, the compute sized from those models' weights, the write paths back into the operator's systems and human control. CodeNinja Atoms has published 23 worked examples of physical AI system design across 9 sectors in the United States (12), Pakistan (6), Saudi Arabia (4), Germany (1), each free to reuse under CC BY 4.0.

## The six decisions in a physical AI system design

| Decision | What the published designs do |
|---|---|
| 1. Sensing and sources | What the system reads: the operator's existing systems of record, read only, plus the cameras, detectors and instruments the operation needs. 22 of 23 designs name their sources on the first page of the paper; every design names them in the body. |
| 2. Where inference runs | 6 of 23 designs state an edge or boundary tier in their At a glance table. By compute pattern: 11 frontier reasoning node, 6 one inference card, 6 no new compute; in every design inference stays in the operator's country. |
| 3. The object model | 317 typed objects and 295 typed links across the 23 designs, each published as hyper-ontology/1 JSON. |
| 4. Open-weight model choice | 57 model choices in all, each with its licence and the reason it was picked; no closed model in any serving path. |
| 5. Sizing from the weights | 23 of 23 designs state their compute and 23 of 23 their three-year cost on the first page, sized from the models' own weights, from no new hardware to an eight-GPU node. |
| 6. Write paths and human control | 23 of 23 designs name their write paths (how a decision is written into the owning system) and 23 of 23 name the person who approves each action that changes the physical world. |

## Worked examples of physical AI system design

| Design | Sector | Country | What it senses or reads | Compute |
|---|---|---|---|---|
| [Wayside Trend Watch](https://codeatoms.ai/wayside-detector-trending-rail-us/) | rail and transportation | United States | An open reference architecture for system design in physical AI. | At the new four-camera yard-entry portal, an octa-core NPU edge box sized from its streams, so only detections, OCR strings and short evidence frames cross a cellular link that drops for hours; centrally, one 141 GB HBM GPU class card in one node in the railroad's own data center, holding the 65.2 GB checkpoint with a 1.2 planning factor and about 63 GB to spare |
| [Warehouse Twin Command](https://codeatoms.ai/warehouse-digital-twin-us/) | warehousing and intralogistics | United States | An open reference architecture for system design in physical AI. | At each site, one industrial edge server with one 48 GB PCIe inference card running vision in Triton and the site model under vLLM; centrally, one 141 GB HBM class node of eight cards, on the operator's own hardware |
| [Pipeline Integrity Watch](https://codeatoms.ai/pipeline-integrity-gas-transmission-us/) | oil and gas | United States | An open reference architecture for system design in physical AI. | At each station, volume production edge modules in 7W to 60W envelopes inside hazardous area enclosures, detecting through backhaul loss; centrally, one node of eight 141 GB HBM-class GPUs in the gas control center |
| [Depot Operating Object Model](https://codeatoms.ai/factory-ontology-rebuild-depot-us/) | discrete manufacturing and automotive | United States | An open reference architecture for system design in physical AI, built on ontologies for unifying factory systems at a government vehicle rebuild depot in the United States, where the inference server that runs it costs about 67,000 US dollars to own over three years. | Fanless DIN-rail edge boxes beside each machine group and at the test cell, their class fixed by a phase one measurement of the vibration load; centrally, two cards of the 48 GB PCIe class air cooled in the depot's own server room on its existing power, holding the 65.2 GB checkpoint with a 1.2 planning factor, about 78.2 GB against roughly 88 GB usable |
| [Port Drone Watch](https://codeatoms.ai/drone-detection-dispatch-port-germany/) | maritime and ports | Germany | An open reference architecture for system design in physical AI, built for a maritime and ports authority in Germany that already flies a docked drone network across two port sites, where owning the inference hardware that runs it costs about 735,000 US dollars over three years. | Three tiers on the authority's own hardware: a GPU-class edge box at each dock, sized at the dock site survey; a site server at each port site with one to four 48 GB PCIe cards; and one node of eight 141 GB HBM cards (1,128 GB) in the port's own data centre, holding 753 GB of FP8 weights with a 1.2 planning factor, 904 GB. Detection events and evidence crops cross the boundary; pixels stay at the dock |
| [Turbine Island Watch](https://codeatoms.ai/steam-turbine-condition-monitoring-pakistan/) | energy and utilities | Pakistan | An open reference architecture for system design in physical AI. | One 48 GB PCIe inference GPU class card on ruggedised edge servers inside the plant perimeter, a PTP grandmaster time card and a one-way egress path; no cloud, because Pakistan has no hyperscaler region |
| [Consist Watch](https://codeatoms.ai/rail-fleet-condition-monitoring-us/) | rail and transportation | United States | An open reference architecture for system design in physical AI. | At the car, an EN 50155 and IEC 60571 module (8 by 8 by 8 inches, IP62, 24 to 48 VDC, CFR 238 materials); centrally, one 141 GB HBM GPU class node in a contractor-hosted US cloud |
| [Hangar Form Cell](https://codeatoms.ai/dieless-sheet-metal-cell-aerospace-mro-us/) | aviation and aerospace manufacturing | United States | a vendor-independent dieless robotic sheet metal cell that takes an aircraft structural part from CAD to a certified, serialized part with collision-free springback-compensated toolpaths, in-process metrology and a per-serial traceability record, wholly inside the facility's own boundary with no vendor cloud. | A fanless industrial inference computer in the cell panel and an eight-card 48 GB PCIe inference server in the facility, joined one way across a hardware data diode; an OCP Time Card grandmaster orders every stream |
| [Corridor Guard](https://codeatoms.ai/dc-interference-gas-pipelines-pakistan/) | oil and gas | Pakistan | a validation-first study that digitises an operator's scanned pipeline records, verifies the corridor in the field, calibrates a direct current interference model in validated software and hands the re-runnable model over on the operator's own hardware, so every safe distance ships with its evidence. | One server-class host of the 48 GB PCIe inference GPU class inside the operator's estate; no field compute, no cloud, nothing crosses the boundary |
| [Pit Camera Watch](https://codeatoms.ai/test-site-camera-monitoring-us/) | discrete manufacturing and automotive | United States | a fully on-premises camera monitoring system that puts every turn, crossing, pit entry and access point of a vehicle proving ground on footage the operator owns, with exportable clips and camera uptime it can prove. | One 48 GB L40S-class card per recorder host, in the conditioned server room; nothing at the poles |
| [Baseline](https://codeatoms.ai/vegetation-mapping-lidar-us/) | agriculture and earth observation | United States | one summer flight of 3-inch 4-band orthoimagery and USGS Quality Level 1 LiDAR, with every tile, point cloud and elevation model bound into a site ontology, so the next season starts from a comparison instead of rediscovery. | The operator's ArcGIS Enterprise; nothing runs in a cloud the operator does not control |
| [Fodder Watch](https://codeatoms.ai/restricted-crop-monitoring-saudi-arabia/) | agriculture and earth observation | Saudi Arabia | an 18-month earth observation service that screens every parcel for restricted green fodder and cultivation beyond the licensed area, checks each detection against the holding's licence, and turns the confirmed ones into case files an inspector can act on. | One 48 GB L40S-class inference node inside the Kingdom; the card class needs a US export licence |
| [Field Ledger](https://codeatoms.ai/farm-data-dashboard-pakistan/) | agriculture and earth observation | Pakistan | one dashboard that joins a farm's live sensor feeds, its historical datasets and a big data and analytics repository into one object model, so any farm, crop cycle or season can be drilled into, exported and reported on under role-based access. | Open-source PostgreSQL with PostGIS on the operator's own servers; source code and full intellectual property handed over |
| [Loop Integrity Watch](https://codeatoms.ai/tank-gauge-integrity-pakistan/) | oil and gas | Pakistan | field hardware that rebuilds the signal path on three Modbus loops, and a governed inventory ontology that catches a flatlined or swapped radar gauge reading before the central inventory picture misleads anyone. | Two containers on the operator's own site application server, inside its OT boundary |
| [Reliability Atlas](https://codeatoms.ai/plant-reliability-assessment-saudi-arabia/) | energy and utilities | Saudi Arabia | a records-based reliability and availability assessment that joins each plant's work orders, trip history and design documents into one reliability object model the operator owns, so every availability figure, criticality rank and root cause finding traces to its source. | No hardware bought; the study runs on the operator's own in-country servers |
| [Grid Context Watch](https://codeatoms.ai/ot-security-cip-evidence-us/) | energy and utilities | United States | one system of context that sees every device on a utility's control networks, flags abnormal behaviour and assembles the evidence a NERC CIP audit asks for, on the operator's own hardware. | One node of eight 141 GB HBM-class GPUs: 753 GB of FP8 weights, 904 GB with the 1.2 planning factor, against 1,128 GB |
| [Structure Phase Watch](https://codeatoms.ai/structure-phase-construction-saudi-arabia/) | heavy industry and construction | Saudi Arabia | one live model of a construction site's structure phase that forecasts schedule slips three days out and shows safety breaches as they happen, on the contractor's own hardware inside the Kingdom. | One node of eight 141 GB HBM-class GPUs holds GLM 5.3 at FP8 (753 GB of weights, 904 GB with headroom) |
| [Steel Count Ledger](https://codeatoms.ai/steel-production-count-pakistan/) | heavy industry and construction | Pakistan | cameras and a GPU industrial PC at every casting strand and cooling bed count billets, ingots, rebars and girders as they are made, publish the counts through a one-way link into one record the authority owns, and let revenue officers reconcile counted against declared production. | One node of eight 141 GB HBM-class GPUs holds GLM 5.3 at FP8 (753 GB of weights, 904 GB with headroom), in a dedicated data center behind an export licence checkpoint |
| [Port Twin](https://codeatoms.ai/port-digital-twin-us/) | maritime and ports | United States | one authoritative digital twin that binds a port's operational systems, live sensor feeds and finance backbone into a thirteen-object ontology, self-hosted on the port's own virtual machines inside the continental United States. | No equipment bought: about 1.1 GB of fp16 weights on the port's existing enterprise virtual machines |
| [Factory Fire Watch](https://codeatoms.ai/factory-fire-monitoring-saudi-arabia/) | heavy industry and construction | Saudi Arabia | live, read-only visibility of fire alarm panels, fire pumps, fire water tanks and energy meters across an operator's highest-risk factories, read through contacts, PLC inputs and LoRaWAN into an IoT platform hosted in Saudi Arabia, with never a write path into certified life-safety equipment. | None bought: the requirement asks for no server or GPU, and the field gateways carry no model runtime |
| [Feeder Firewatch](https://codeatoms.ai/wildfire-risk-distribution-us/) | energy and utilities | United States | a live ignition and outage risk score for every distribution feeder segment, built on the cooperative's own hardware, with every de-energisation and fast-trip change approved by a named operator. | One node of eight 141 GB HBM-class GPUs holds GLM 5.2 at FP8 (753 GB of weights, 904 GB with headroom) |
| [Terminal Pulse](https://codeatoms.ai/truck-turn-container-terminal-us/) | maritime and ports | United States | truck turn time predicted two hours out and yard congestion seen live, at a container terminal, on the operator's own hardware with no outbound connection. | One node of eight 141 GB HBM-class GPUs holds GLM 5.3 at FP8 (753 GB of weights, 904 GB with headroom), at about 10 kW |
| [Sovereign HSE Watch](https://codeatoms.ai/sovereign-hse-pakistan/) | oil and gas | Pakistan | An open reference architecture for predicting HSE incidents at an oil and gas operator in Pakistan, on the operator's own hardware, with no data leaving the country and no third-party AI service in the serving path. | One node of eight 141 GB HBM-class GPUs holds GLM 5.3 at FP8 (753 GB of weights, 904 GB with headroom) |

## Where to learn physical AI system design

Design on Praxis: https://codeatoms.ai/praxis/ · PADI, the benchmark for physical AI system design: https://codeatoms.ai/padi/ · Developers: https://codeatoms.ai/developer/ · Dataset: https://huggingface.co/datasets/CodeNinjatools/vertical-driven-architectures

See also: [sovereign AI reference architectures](https://codeatoms.ai/sovereign-ai-reference-architectures/), [industrial ontology examples](https://codeatoms.ai/industrial-ontology-examples/), [industrial AI reference architecture](https://codeatoms.ai/industrial-ai-reference-architecture/), [self-improving AI system](https://codeatoms.ai/self-improving-ai-system/).