DEV Community: Umar Bilal

System design for physical AI: steam turbine condition monitoring that can never write to the DCS

Umar Bilal — Sun, 25 Oct 2026 05:14:27 +0000

This is the engineering summary of an open reference architecture. The paper, its object model as JSON and the model register are free to reuse under CC BY 4.0: https://codeatoms.ai/steam-turbine-condition-monitoring-pakistan/. DOI 10.5281/zenodo.23365092. The operator is described by class, never by name.

A combined cycle plant in Pakistan is rehabilitating a steam turbine and installing a new 280 MW, 20 kV hydrogen-cooled generator. The protection layer already trips the unit when bearing vibration reaches Zone C. The question the design answers is the one before the trip: can the shift operator see a developing excursion, with a prescribed action, while the bearing housing is still drifting up from its nominal 2.3 mm/s RMS?

1. Read the DCS, never write to it

The DCS is read through a hardware one-way egress device into a DMZ OPC UA collector. The platform gets process, alarm and vibration context and has no path back into the control network, so the existing trip philosophy is untouched. The design chose this over a direct platform connection into the control network because the NEPRA Security of IT and OT Regulations 2022 and the sector's own pitfalls rule out a live bidirectional path for an analytics layer. A signed zone and conduit diagram is a phase one deliverable, not an afterthought.

2. Eight sources, four adapter families, one model

Eight named systems feed the platform: the DCS, the 3500 vibration monitoring rack, the generator condition monitoring systems, the document management system, the programme schedule, the plant CMMS or EAM, SCADA and outage management. Four adapter families (integration, file and document, operational systems, telemetry) turn each source's protocol and naming into instances of one ten-object model, so nothing downstream knows that a vibration channel arrived as an OPC UA node and a test report arrived as a document. The plant CMMS is marked domain typical: it carries no commitment until the operator confirms it.

3. Ten objects, typed links

The condition model holds the steam turbine, the new generator, the vibration rack, the condition anomaly alert, the condition work order, the instrument calibration record, the commissioning and performance test record, the shift operator, the warranty engineer and an availability measure. The links are typed, so one traversal answers "every vibration anomaly since the rehabilitation, the calibration state of each channel when it fired, and the work order and warranty engineer now attached". In a document store that question is application code that joins loose records by hand.

4. Two open-weight models, no training run

MOMENT-1 (MIT) runs zero-shot anomaly detection over vibration, hydrogen, winding and process telemetry, because the rehabilitated unit's recovered history may be too sparse to baseline every channel. IBM Granite 4.2 8B (Apache-2.0) drafts diagnostic notes and test reports for engineer review; it was chosen over a 27B model as the smaller, licence-clean fit with real KV headroom on one card. Both run on one 48 GB PCIe inference class card on ruggedised edge servers inside the plant perimeter. Pakistan has no hyperscaler region, so no tier assumes a cloud, and model updates arrive by mirror-archive delivery on the site's own import cycle.

5. What it costs

Appendix A prices the site envelope (the host share for one L40S-class card, a PTP grandmaster time card, two industrial switches and a UPS) at 21,600 to 23,400 US dollars over three years, about the same as the deepest three-year AWS commitment for one L40S card in the UAE region, which could not keep the read one-way in any case. The data diode and enclosure are priced by quotation and left out.

6. The gate that matters

Alert flooding trains operators to ignore alerts, so the design carries a gate of 2 false alerts per week, curated unit-checked series, and channel health separated from process alarms. The platform advises; the protection layer trips.

Designed on CodeNinja Praxis, the CodeNinja Atoms platform for designing physical AI systems. The object model imports into Hyper Ontology. Paper, ontology and register: https://codeatoms.ai/steam-turbine-condition-monitoring-pakistan/

One of the CodeNinja Atoms sovereign AI reference architectures: every published design with its sector, country, compute and three-year cost.

System design for physical AI: a dieless robotic sheet metal cell that owns its toolpaths, weights and evidence

Umar Bilal — Sat, 24 Oct 2026 03:03:26 +0000

This is the engineering summary of an open reference architecture. The paper, its object model as JSON and the model register are free to reuse under CC BY 4.0: https://codeatoms.ai/dieless-sheet-metal-cell-aerospace-mro-us/. The operator is described by class, never by name.

An aircraft maintenance, repair and overhaul facility in the United States needs structural brackets, panels and skin sections on demand. Today a low volume part waits for a die that must be designed, cut, stored and paid for before the first blank is bent, or waits for a supply chain the facility does not control. The dieless robotic alternatives on the market arrive with hosted reachback services the facility cannot own, which flight critical provenance rules will not tolerate. The question is narrow: can this serial numbered part be formed now, on a machine the facility owns, to drawing tolerance, with evidence a quality authority will accept, without a die and without a vendor's cloud?

1. Four constraints before any component is named

No vendor cloud and no reachback, because camera imagery and process data on controlled parts are controlled technical data. The facility owns every artifact: weights, trained detectors, toolpath baselines, the record layer and every container image. The record attaches to the serialized part instance, never the part type. And the riskiest assumption is tested first, in metal, on sacrificial aluminium, with a written kill criterion in the facility's own numbers. Each constraint eliminates whole families of choices before anything is bought.

2. One joined record per serialized part

Fourteen objects held as Asset Administration Shells in Eclipse BaSyx: the formed part, the sheet blank with its mill certificate and lot, the dieless cell, the robot arm and stylus, the toolpath program with its CAD revision and compensation passes, the CAD model, the formability analysis, the metrology scan with its instrument calibration state, the force trace, the work order, the nonconformance with its evidence pack, the facility operator and the cell model service with its weights hash. From one serial number a query reaches the mill certificate, the CAD revision, every scan and the calibration state of the instrument that took it, every force trace, the operator who released the program and any disposition. The first article inspection pack is assembled by traversal, not by hand.

3. Six sources, two adapter families, one clock

CAD import, the operator's own sensors, third party software interfaces, the production systems, the quality records system and cell control telemetry over MTConnect all enter read only through an integration adapter or a telemetry adapter; nothing reaches the model directly. Apache Kafka 4.3 in KRaft mode, self managed inside the perimeter, orders events by part serial. An OCP Time Card grandmaster with linuxptp keeps a scan, its force trace and the arm pose on one clock, because a joined record without a shared clock is a guess with a timestamp attached.

4. Detection at the cell, governance in the facility

A fanless industrial inference computer in the cell panel serves Anomalib 2.7.0 for surface and geometric anomaly detection on in-process and post form scans, and Chronos-Bolt for force trace baselines, both Apache 2.0, through ONNX Runtime, interleaved between passes so a flagged pass stops before the next begins. An eight card 48 GB PCIe server in the facility holds the registered models, the thresholds and the versioned springback baselines. Records cross from the cell's OT zone to the facility tier one way, over a hardware data diode running lidi and Eurydice; released programs reach the cell by a signed release path, never a return channel. No language model is committed, so no frontier cluster is sized.

5. The safety function admits no software verdict

Light curtains, the safety PLC and the interlocks own stopping. Around them sit three advisory budgets: detection inside the pause between passes, scan evaluation between acquisition and the deviation map reaching the record, alert routing while the part is still on the fixture. Each is verified against bench measured streams, not asserted on paper. The anomaly detector runs in shadow until its false alarm rate on the facility's own panels is known; only then does its verdict hold a part.

6. Three phases, one kill criterion

Twelve items in a validation prototype on sacrificial blanks, where dual arm synchronization on a large flexible panel, the assumption a previous industry fuselage robotics program failed on, is proven or the work stops at prototype cost. Five items for the production cell, exiting on facility staff running it on their own authority. One item for a steel and titanium limits investigation, reporting achievable accuracy as a range before any production claim.

The design is one of the sovereign AI reference architectures CodeNinja Atoms publishes, each with its sector, country, compute and cost in one table.

The cost line

The priced classes come to about 151,000 US dollars in list prices, and 202,000 to 220,000 over three years with support and power, the structured light scanner and the eight card server most of it. The two compute tiers alone cost about 132,000 dollars to own over three years, about two fifths of the deepest three year AWS commitment for eight L40S cards, which the offline requirement rules out anyway. The laser tracker, the data diode and the safety PLC are sold by quotation and are named but not totalled. The sourced arithmetic is in Appendix A.

Designed on CodeNinja Praxis, the CodeNinja Atoms platform for designing physical AI systems. The object model imports into Hyper Ontology. DOI: 10.5281/zenodo.23308095.

System design for physical AI: prove every pipeline record before the interference model reads it

Umar Bilal — Fri, 23 Oct 2026 02:43:12 +0000

This is the engineering summary of an open reference architecture. The paper, its object model as JSON and the model register are free to reuse under CC BY 4.0: https://codeatoms.ai/dc-interference-gas-pipelines-pakistan/. The operator is described by class, never by name.

A gas pipeline operator in Pakistan has two buried lines within a kilometre of a high voltage direct current electrode station. When the HVDC link runs in monopolar mode, up to 1.3 kA returns through the ground, and some of it rides the operator's steel. The question the operator cannot answer today: along which sections does that current push pipe to soil potential past the cathodic protection criterion of minus 0.85 V against a copper sulfate electrode, and what mitigation, with what bill of quantities, restores it?

The records that should answer it are scanned survey sheets, coating reports, layouts and soil investigations with no provenance on any value, and the tender quotes a single backfill resistivity for ground the same records describe as layered and variable. The contract makes every data error the contractor's cost and pays only after the final report. That shapes the whole design.

1. The register is the product, not the model

The first deliverable is a validation register: fifteen typed objects (the two lines and their insulating joints, the anode beds and the generator unit, the electrode station, the HVDC system with its operating states, the test points, the soil report, the assessment and each mitigation measure), with every attribute labelled measured, inherited, assumed or unknown. Nothing enters it without a label. A safe distance computed from an assumed resistivity cannot be defended; a safe distance whose inputs trace to a dated field reading can.

2. One parser, two keyers, one engineer

The only neural workload is PaddleOCR-VL 1.6, about 0.9 billion parameters under Apache 2.0, digitising the scanned package. It runs as released, never fine tuned on operator data, so the weights handed over are the released weights and no training data obligation exists. Every extracted value is captured by two independent keyers; a disagreement becomes a logged discrepancy, and every discrepancy routes to the operator's corrosion control engineer before any modelling starts. No open weight model ever touches a safety calculation.

3. The physics stays in validated software

The interference model is built in the commercial electromagnetic and grounding software the governing clauses require (CDEGS or COMSOL class). The design calibrates it against bipolar baseline field readings, because the monopolar returns that drive the worst case are rare and short, and models the monopolar and fault cases as a conservative envelope with the validation basis and its limits written into the report. Layered soil enters the model as separate regions, never as one bulk value.

4. What the stack leaves out

No streaming backbone, no historian, no video ingest, no orchestrator, no edge compute, no language model. The study consumes a furnished document package, nothing moves, nothing is tracked, and the deliverable is an engineering report authored in the mandated software. One server class host of the 48 GB PCIe inference class inside the operator's estate carries the parser and the solver runs; parser weights take under 2 GB of the card and the rest is solver memory. Nothing crosses the boundary in either direction, and the re-runnable model is handed over on the operator's own hardware with no consultant side copy surviving.

5. Gates spend the data risk before the clock does

Three phases carry fifteen items. Phase one exits only when the register is complete with provenance on every attribute, every discrepancy is closed by the named engineer and the corridor within 1000 m of the station is field verified. Phase two exits on a calibrated model and an accepted safe distance assessment. Phase three exits on an accepted mitigation strategy with its bill of quantities and the model handed over. A data problem found after the final report has no schedule or payment recourse at all, which is why the gates are dated and front loaded.

The cost line

Owning the one host costs about 19,500 US dollars over three years, inside 18,700 to 20,200 with support and power. Renting one L40S card from the nearest AWS region would cost 22,500 to 60,000 dollars, so ownership is about four fifths of the deepest three year commitment, and the rental could not keep the data in Pakistan in any case. The sourced arithmetic is in Appendix A of the paper.

This design is listed with every other published design among the sovereign AI reference architectures at CodeNinja Atoms.

Designed on CodeNinja Praxis, the CodeNinja Atoms platform for designing physical AI systems. The object model imports into Hyper Ontology. DOI: 10.5281/zenodo.23308093.

System design for physical AI: a camera count is not coverage, so prove it pole by pole on a vehicle proving ground

Umar Bilal — Thu, 22 Oct 2026 02:41:59 +0000

This is the engineering summary of an open reference architecture. The paper, its object model as JSON and the model register are free to reuse under CC BY 4.0: https://codeatoms.ai/test-site-camera-monitoring-us/. The operator is described by class, never by name.

A vehicle proving ground in the United States runs test vehicles through marked turns, a pit entry and crossings where people walk. Its camera system was bought in 2013 and has decayed into outages, dead heads and no redundancy, and nobody has a diagram of the cable routes. The question it cannot answer: where was every vehicle and every person when something happened, and can you show it again thirty days later?

The tempting answer is a purchase order with a camera count on it. Here is why that is the wrong unit, and what the design does instead.

1. Survey first, because a camera count is not coverage

No pole is ordered until a coverage survey shows every crossing, pit entry, turn and access point held at the required pixels on target, with low light treated as a budget and mounting geometry deciding what each head actually sees. Any dead spot is closed by geometry or accepted in writing. The survey evidence pack is the phase one exit gate, so the money at risk is still small if it fails.

2. Two strands per pole, home to one room

40 to 60 poles carry fixed thermal and visible heads, no wireless radios, each fed by Power over Ethernet from managed switches, with NEMA rated enclosures and step-down transformer feeds at the poles. New underground single-mode fiber runs two strands to every pole and lands at the operator's own termination building, where the recording tier sits in a conditioned server room. Every head is screened under Section 889 with a component-level bill of materials before purchase, and an adapter rejects any device it was not provisioned for.

3. An N+1 recorder you own outright

Open-source Frigate NVR records 80 to 120 streams on a primary and a standby host, each holding the full 30 days on local disk. There is no message broker, because nothing downstream subscribes; ordering comes from time instead, an OCP Time Card grandmaster with a holdover oscillator disciplining every host, so an exported clip is evidentiary on its timestamps alone.

4. One detector, searched after the fact

The estate is passive: people watch, nothing raises an alarm, so no model sits in the live path. D-FINE, Apache-2.0 on both code and weights, runs over the archive when the safety office asks a question, such as every clip of a person near a turning vehicle. Its weights are about 8 MB for the N tier and 124 MB for X, a rounding error on a 48 GB card whose real budget is decode. RF-DETR was set aside because its XL tiers sit under a platform license.

5. What it costs, and what it never costs

Owning the whole estate costs about 474,000 to 871,000 US dollars over three years from public list prices, and the camera heads are most of it: a thermal and visible head is seven to eight thousand dollars, one per pole. The compute is the small line. Two recorder hosts with one L40S class card each cost about 32,600 dollars over three years, about four fifths of renting the same two cards on the deepest three-year AWS commitment, and that rented figure stores none of the footage. What it never costs is a license per camera: the recorder is open source and the detector is Apache-2.0, so adding the sixty-first pole adds hardware and nothing else.

6. Acceptance per use case, not per camera

Camera uptime is a first-class metric with re-runnable quarterly acceptance tests, and the estate is accepted one use case at a time: a pedestrian crossing watched through a shift, a test cycle reconstructed from 30 days of footage. The last gate is forensic search returning the right clip from the right camera at the right time.

Full design, figures, the cost appendix with every price cited, and the fifteen-object model: the paper. DOI: https://doi.org/10.5281/zenodo.23296077

Its fifteen objects sit beside every other published object model in the industrial ontology examples collection.

Designed on CodeNinja Praxis, the CodeNinja Atoms platform for designing physical AI systems. The object model imports into Hyper Ontology, which turns it into a living system. Load it yourself with the open hyper-ontology loader.

Physical AI System Design for Warehousing and Intralogistics: What the Published Designs Already Answer

Umar Bilal — Wed, 21 Oct 2026 02:52:26 +0000

First published on CodeNinja Atoms.

A warehouse asks the same first questions whether it is a distribution center in the United States or a cross dock in Saudi Arabia: what is happening across the building right now in one picture, why the trucks at the dock are waiting, who is walking where the forklifts drive, which conveyor or sorter drive is about to stop, is the fire protection actually armed, and how much left the building against what the system says shipped. The series on this site holds no warehousing design yet. It holds eighteen open reference architectures in other sectors, and nine of them are written for an operation that asks one of those questions in another setting. This guide reads them as a warehouse operator would, using only what each paper says.

The map

The warehouse's question	The published design that answers it	Where it was written for
One picture of the operation from the systems it already runs	Port Twin	A landlord port authority in the United States
Why the trucks at the dock and yard are waiting	Terminal Pulse and Structure Phase Watch	A container terminal in the United States; a construction site in Saudi Arabia
Who is near the forklifts and yard vehicles	Pit Camera Watch and Terminal Pulse	A vehicle proving ground in the United States; a container terminal in the United States
Which vehicle in the fleet is raising a fault	Consist Watch	A 225 car commuter rail fleet in the United States
Which conveyor or sorter drive is about to fail	Turbine Island Watch	A combined cycle power plant in Pakistan
Is the fire protection armed	Factory Fire Watch	High risk factories in Saudi Arabia
How much product left the building	Steel Count Ledger	Steel mills in Pakistan, counted for a public authority
What the records already prove	Reliability Atlas	Desalination and treatment plants in Saudi Arabia

One picture from the systems already running

Port Twin is the design to read when the problem is that every dashboard is a slice. It binds eight operational systems, five live sensor feeds and one finance backbone into a thirteen object model with the berth as the focal object, joined through three adapter families and one event backbone. The rule a warehouse will recognise is the boundary: the twin shows, and pilots, berth planners, engineers and finance staff decide in their own systems, with nothing written back into a system of record. The only model in the register is one self hosted open weight embedding model, BGE-M3, for document retrieval, about 1.1 GB of weights on virtual machines the port already runs, so no equipment is bought. For a distribution center, the warehouse management system, the transport management system, the yard system and the building's sensors take the places of the port's systems, and the dock door takes the place of the berth; the pattern of a read only twin over owned systems carries across unchanged.

Dock and yard turn time

Terminal Pulse predicts truck turn time two hours out at a container terminal and watches the yard live. The paper's own numbers frame the problem a dock office knows: turn time averages 54 minutes and spikes above 90, and the causes live in separate systems, so planners stitch them together at the morning and afternoon operations meetings and a spike is explained hours after the trucks felt it. The design joins eleven systems, among them the terminal operating system, the gate system, two crane management systems, reefer monitoring, railroad switch lists, cameras and weather and tide feeds, into twelve objects that include the truck visit, the gate lane, the yard block and the transfer zone. Five self hosted open models carry the work: GLM 5.3 for reasoning, Chronos-2 for forecasting, Qwen3-Embedding-0.6B for retrieval, and RF-DETR with Roboflow trackers for vision. Two choices transfer directly to a dock. The first rollout gate proves or redirects the yard side hypothesis using gate OCR data the operator already holds, before anything else is built. And the surfaces warn and propose while the named planner acts inside the terminal operating system; write back into that system is a second phase option gated on labour and IT sign off. The owned stack is costed at about 722,000 US dollars over three years. The operator in this paper is an illustrative scenario, not a customer.

Structure Phase Watch adds the inbound side: gate movements and delivery windows are objects in its model, read from haulage GPS, gate logs and biometric turnstiles, and every delivery window it drafts is a recommendation recorded with the approver. Its open item of 13 October is the one any camera estate on a loading apron will meet: the paper names dust, glare and heat as what degrades the lenses, and the answer recorded on its page is a cleaning rota plus a confidence drop on a fixed reference target in each camera's view that opens a work order for a named person, with the camera's health held in the object model as a record someone owns.

People and vehicles on camera

Pit Camera Watch is a fully on premises camera system for a vehicle proving ground: 40 to 60 poles of thermal and visible camera heads, 80 to 120 streams, recorded on an N+1 pair of open source recorders holding 30 days of footage on local disk inside an isolated network. Its human rule is the conservative one a warehouse safety office may want first. Booth operators watch and decide, nothing raises an automated alarm, and only the safety office can export a clip. The one model, D-FINE under Apache-2.0 at about 8 to 124 MB of weights, searches the archive after the fact for a test vehicle, a support vehicle or a pedestrian; it is not in the live safety path. For aisles and docks, the same shape gives an investigation a searchable archive the operator owns without putting a model between a person and a near miss.

Terminal Pulse is the live counterpart. Its model holds a yard person, a transfer zone and a safety event, its detectors and trackers run in fanless enclosures at the yard blocks, gate and quay, no safety reflex crosses a network hop, and the named safety supervisor acknowledges every safety event. A warehouse choosing between the two is choosing whether the cameras search the past or watch the present, and both papers show what each choice costs in hardware and in human workload.

Fleet faults: the operator's own logic, executed on the vehicle

Consist Watch monitors 225 commuter rail cars, and the part a forklift or tugger fleet can take is where the logic runs. An onboard module reads 64 named signals per car, 32 digital and 32 analog plus an accelerometer and GPS, and executes the operator's own fault logic table verbatim at the car. Telemetry rides the vehicle's existing cellular routers, a named mechanical desk controller owns every active fault in that territory, and alerts reach named recipient groups by severity. No training run was needed: Chronos-2 handles analog anomaly detection and seasonal trending, and a self hosted open weight reasoning model carries the work surface over the fleet ontology. A truck maker already publishes the fault codes for its vehicles; this design says where to execute them and who owns the result.

Conveyors and sortation: advise, never trip

Turbine Island Watch reads a rehabilitated steam turbine generator through a one way path from the control system, the vibration monitoring rack and the generator protection systems. Three of its rules transfer to a conveyor drive, a sorter or a stacker crane without change. The read is one way and nothing writes to the control system. The platform advises and the protection layer trips, so every alert carries a prescribed action that a named shift operator acknowledges. And the rollout gate is a number: two false alerts per week, with thresholds tuned only on channels that have provable history. MOMENT-1 detects anomalies zero shot and no training run is needed. The compute is one 48 GB inference card on ruggedised edge servers inside the plant perimeter, costed at about 22,500 US dollars to own over three years.

Fire and life safety, read only

Factory Fire Watch gives live visibility of fire alarm panels, fire pumps, fire water tanks and energy meters across an operator's highest risk sites, read through panel contacts, PLC inputs and LoRaWAN into an IoT platform hosted in the Kingdom, with never a write path into certified life safety equipment. No server or GPU is bought. The field equipment per site is about 2,600 to 3,200 US dollars in list prices, and a monitoring officer acknowledges or escalates every safety critical alert. The paper's framing suits a high bay building well: a sensor that has stopped reporting still reads as normal, so the design treats silence as a failure to be surfaced. For a distribution center and its sprinkler pumps, the design is the same read only estate with the sites renamed.

Counting what leaves the building

Steel Count Ledger puts cameras and one sealed GPU industrial PC at every installation point, counts product as it is made, and publishes the counts through a one way link into one record the authority owns. Counting survives a slow wide area link because it happens at the edge, tracker models keep one item from counting twice, and the weighbridge feed serves as acceptance ground truth. Every discrepancy case is judged by a named revenue field officer; the system counts and reconciles and never assesses, and the mill never edits a count. The same pattern counts pallets or cases at an outbound dock door against what the warehouse management system says shipped, with the detector fine tuned per product type and site as the paper describes.

What the records already prove

Reliability Atlas is the design to read before any sensor is bought. It joins each plant's work orders, trip history and design documents through three read only adapters, as exports rather than live interfaces, into a reliability object model the operator owns. Availability is computed by Monte Carlo simulation on reliability block diagrams built from each plant's own failure study, auditable to ISO 55000, with predictions accepted only by a named reviewer and root cause reports validated only by a named engineer. The model register is empty by design, no hardware is bought, and the study runs on the operator's own in country servers. A warehouse with a maintenance system and years of conveyor and sorter downtime records can run this study first and let it decide which of the designs above it needs second.

What every one of them keeps constant

Across all nine, four things do not vary. The system runs on hardware the operator controls, or in a cloud or region the operator chose. A named person owns every decision, and nothing executes on equipment. The object model is published as JSON in the hyper-ontology/1 format, so a warehouse can read the exact objects and links before building any of its own. And the cost is computed the same way in every appendix, from the paper's own sizing with every price cited, so the own or rent question has one answer per design rather than a slogan; the comparison across designs is in Own or Rent the GPUs.

What a first warehousing design would have to name

None of the nine holds a storage location, a pick face, an order line, a dock door, a forklift or a pedestrian aisle as an object, and none holds slotting: which item belongs in which location, and why. Terminal Pulse holds a yard block and a container, which is the nearest shape, but no published design reasons over where stock should sit. A first warehousing design adds that object model, and the near miss between a person and a vehicle in an aisle with it. It is the design the co-build page for warehousing and intralogistics is open for: the operator brings the operation, the systems that already hold its records and one engineer who knows the building; the published designs above are the proof of what comes out.

Use the designs

Every object model named here is in the paper's package as ontology/objects.json, and the same rows are in the Vertical-Driven Architectures dataset and served live by the codeninja-research MCP server, so a coding agent can ask which published design holds a truck visit or a safety event object and read why. Each paper is citable by its DOI on its page. All were reasoned on CodeNinja Praxis and their object models import into Hyper Ontology.

Every design cited here is one row in the table of sovereign AI reference architectures, and every object model is browsable among the industrial ontology examples.

About CodeNinja

CodeNinja is a Middle Eastern-American artificial intelligence lab focused on building self-improving systems. We are reinventing knowledge work to close the loop between vertical AI use cases and the generalized intelligence that fuels it, accelerating the path toward organizational superintelligence.

Physical AI System Design for Mining: What the Published Designs Already Answer

Umar Bilal — Tue, 20 Oct 2026 02:39:46 +0000

First published on CodeNinja Atoms.

A mine asks the same first questions whether it is an open pit in Saudi Arabia or an underground operation in the United States: who is near a haul truck right now, which machine is about to fail, which shift is drifting toward an incident, is the fire protection at the plant actually armed, and how much ore left the gate against what was declared. The series on this site holds no mining design yet. It holds eighteen open reference architectures in eight other sectors, and eight of them are written for an operation that asks one of those questions in another setting. This guide reads them as a mining operator would, using only what each paper says.

The map

The mine's question	The published design that answers it	Where it was written for
Who is near the vehicles on the haul road	Pit Camera Watch and Structure Phase Watch	A vehicle proving ground in the United States; a construction site in Saudi Arabia
Which mobile machine is about to fail	Consist Watch	A 225-car commuter rail fleet in the United States
Which fixed plant is about to fail	Turbine Island Watch	A combined cycle power plant in Pakistan
Which shift is drifting toward an incident	Sovereign HSE Watch	An oil and gas operator in Pakistan
Is the fire protection armed	Factory Fire Watch	High-risk factories across industrial cities in Saudi Arabia
How much product left the gate	Steel Count Ledger	Steel mills in Pakistan, counted for a public authority
What the records already prove	Reliability Atlas	Desalination and treatment plants in Saudi Arabia

People and vehicles on the haul road

Pit Camera Watch is a fully on-premises camera system for a vehicle proving ground: 40 to 60 poles of fixed thermal and visible camera heads on new underground fiber, 80 to 120 streams, recorded on an N+1 pair of open-source recorders holding 30 days of footage on local disk inside an isolated network. The design's human rule is the one a mine's safety office will recognise. Booth operators watch and decide, nothing raises an automated alarm, and only the safety office can export a clip. The one model in the register, an Apache-2.0 detector of about 8 to 124 MB, searches the archive after the fact for a test vehicle, a support vehicle or a pedestrian; it does not sit in the live safety path. For a haul road, the same shape holds: the footage is the operator's, the retrospective search is what an investigation runs, and the decision about a near miss stays with a person.

Structure Phase Watch adds the part a pit needs and a proving ground does not: edge compute in the field. Its five solar powered edge nodes sit at the gate, the laydown area, the crane slew zone, the batch plant and the haul road, and every breach alert is confirmed, dismissed or escalated by the HSE officer. Its open item of 13 October is the most mining-relevant sentence in the series: the paper names dust, glare and heat as what degrades the camera lenses, and the answer recorded on the paper's page is both a cleaning rota set by the site's dust season and a confidence drop on a fixed reference target in each camera's view that opens a work order for a named person, with the camera's health held in the object model as a record someone owns.

Mobile fleet: the operator's own fault logic, executed at the machine

Consist Watch monitors 225 commuter rail cars. The part a mining fleet can take verbatim is where the logic runs. An onboard module reads 64 named signals per car, 32 digital and 32 analog plus an accelerometer and GPS, and executes the operator's own fault-logic table at the car. Telemetry rides the vehicle's existing cellular routers, and severity-ranked faults reach a named mechanical desk controller who owns every active fault in that territory, with alerts routed by severity to named recipient groups. The module is specified to the rail standards for shock, vibration and temperature. A haul truck's manufacturer already publishes the fault codes; this design says where to execute them, who owns the result, and that no training run was needed: an open-weight time series model handles analog anomaly detection and seasonal trending, and a self-hosted open-weight reasoning model carries the agentic work surface over the fleet ontology.

Fixed plant: advise, never trip

Turbine Island Watch reads a rehabilitated steam turbine generator through a one-way path from the distributed control system, the vibration monitoring rack and the generator protection systems. Three of its rules transfer to a mill, a crusher or a conveyor drive without change. The read is one-way and nothing writes back to the control system. The platform advises and the protection layer trips, so every alert carries a prescribed action that a named shift operator acknowledges. And the rollout gate is a number: two false alerts per week, with thresholds tuned only on channels that have provable history. The compute is one 48 GB inference card on ruggedised edge servers inside the plant perimeter, costed at about 22,500 US dollars to own over three years.

Incidents before they happen

Sovereign HSE Watch predicts health, safety and environment incidents at an oil and gas operator by joining eight systems: the enterprise HSE, maintenance and quality modules, SCADA, fire and gas, vision cameras, SQL databases, scanned files and the identity provider. Every recommendation is approved or rejected by a named person and nothing executes on equipment. It runs six self-hosted open-weight models on one node of eight 141 GB HBM-class GPUs, and its first rollout gate is not a model but a licence: 141 GB-class accelerators need a US export licence for Pakistan, so the rollout confirms installed hardware before anything else. A mine in the same country group inherits that gate unchanged, and a mine in Saudi Arabia reads the Structure Phase Watch appendix for the equivalent arithmetic.

Fire and life safety, read only

Factory Fire Watch gives live visibility of fire alarm panels, fire pumps, fire water tanks and energy meters across an operator's highest-risk sites, read through panel contacts, PLC inputs and LoRaWAN into an IoT platform hosted in the Kingdom, with never a write path into certified life-safety equipment. No server or GPU is bought. The field equipment per site is about 2,600 to 3,200 US dollars in list prices, and a monitoring officer acknowledges or escalates every safety-critical alert. For a processing plant and a camp, the design is the same read-only estate with the sites renamed.

Counting what leaves the plant

Steel Count Ledger puts cameras and one sealed GPU industrial PC at every casting strand and cooling bed, counts product as it is made, and publishes the counts through a one-way link into one record the authority owns. Counting survives a slow wide-area link because it happens at the edge. Every discrepancy case is judged by a named revenue field officer; the system counts and reconciles and never assesses, and the mill never edits a count. The same pattern counts trucks at a weighbridge or product at a load-out, with the detector fine-tuned per product type and site as the paper describes.

What the records already prove

Reliability Atlas is the design to read before any sensor is bought. It joins each plant's work orders, trip history and design documents through three read-only adapters, as exports rather than live interfaces, into a reliability object model the operator owns. Availability is computed by Monte Carlo simulation on reliability block diagrams built from each plant's own failure study, auditable to ISO 55000, with predictions accepted only by a named reviewer and root cause reports validated only by a named engineer. The model register is empty by design, no hardware is bought, and the study runs on the operator's own in-country servers. A mine with a maintenance system and a historian can run this study first and let it decide which of the designs above it needs second.

What every one of them keeps constant

Across all eight, four things do not vary. The system runs on hardware the operator controls, or in a cloud the operator chose in its own country; the papers for Pakistan say plainly that no hyperscaler runs a region there. A named person owns every decision, and nothing executes on equipment. The object model is published as JSON in the hyper-ontology/1 format, so a mine can read the exact objects and links before building any of its own. And the cost is computed the same way in every appendix, from the paper's own sizing with every price cited, so the own or rent question has one answer per design rather than a slogan; the comparison across designs is in Own or Rent the GPUs.

What a first mining design would have to name

None of the eight holds a haul truck, a light vehicle, a haul road segment or a pedestrian exclusion zone as an object, and none holds a near miss between them. That is the object model a first mining design adds, and it is the design the co-build page for mining is open for: the operator brings the operation, the systems that already hold its records and one engineer who knows the site; the published designs above are the proof of what comes out.

Use the designs

Every object model named here is in the paper's package as ontology/objects.json, and the same rows are in the Vertical-Driven Architectures dataset and served live by the codeninja-research MCP server, so a coding agent can ask which published design holds a monitoring module or a fault history object and read why. Each paper is citable by its DOI on its page. All were reasoned on CodeNinja Praxis and their object models import into Hyper Ontology.

Every design cited here is one row in the table of sovereign AI reference architectures, and every object model is browsable among the industrial ontology examples.

About CodeNinja

What Does It Take to Vibe Code a Truck?

Umar Bilal — Mon, 19 Oct 2026 02:39:43 +0000

First published on CodeNinja Atoms. The designs it draws on: sovereign AI reference architectures.

A year ago, vibe coding meant something new. You describe the app you want, a model writes it, you run it, you describe what is wrong, and it fixes it. A person with no training in software can now ship working software.

So here is the question I keep coming back to. What would it take to vibe code a truck?

Not the cab or the paint. The whole system. An autonomous haul truck working in a copper mine, the fleet it belongs to, the network it talks over, the people who work around it, and the regulator who has to sign it off.

Today no model can do that. This post is about why, and about what we are building at CodeNinja Atoms to change it.

The Truck

Start with what an engineer actually has to decide before an autonomous haul truck moves a single tonne.

Sensing. Which lidar, radar and cameras, where they sit, and what they can see in dust, rain and at night. A sensor that works on a highway can fail on a haul road.

Compute. What runs on the truck, what runs at the edge of the pit, and what runs in a data center. How much power, cooling and memory each one needs, and what happens when the link drops.

Models. Which perception, prediction and planning models, at what size, trained or adapted on what data, and how each one is monitored once it is running.

The site. Haul roads, loading points, crushers, fuel bays, light vehicles driven by people, and the private network that ties them together. The truck is only safe if the site is designed around it.

Safety and approvals. Autonomous mining machines have their own safety standards, and each country adds its own rules on top. Someone has to write the safety case, and someone has to approve it.

Cost and time. What the system costs over three years, what it saves, and how long it takes before the first truck runs on its own.

Maintenance. How the fleet knows a part is about to fail, and who acts on it.

Every one of these decisions depends on the others. Change the sensor and the compute changes. Change the country and the approval path changes. That web of decisions is system design, and for physical AI it is one of the hardest problems there is.

The Bottleneck

A model can write software because it has read more software than any person ever will. Billions of lines of code are public, along with the questions people asked about them and the answers that worked.

Physical system design is not like that. The knowledge sits in tender documents, internal standards, site reports and, above all, in the heads of engineers who spent twenty years in mining, oil and gas, ports or power. Very little of it is written down, and almost none of it is public.

So the bottleneck in physical AI is not the model. It is deep domain expertise. A young engineer can learn to code in a year. Learning to design a mine takes a career.

That is the problem we set out to solve. Our mission is to let a young engineer design the physical AI systems that whole industries run on, without first spending twenty years in those industries.

CodeNinja Atoms

CodeNinja Atoms is how we work on that mission in the open. It has four parts.

Praxis is the platform. An engineer brings a requirement, often a real tender, and Praxis reasons out the full system design: sensing, compute, models, the object model of the operation, cost, timeline and the approvals it will need. It explains every choice, so the engineer learns the design instead of just receiving it. Praxis is free for engineers.

Hyper Ontology is the shared language of those designs. Every design carries an object model of the operation, the assets, people, events and rules, and how they connect, in an open JSON format called hyper-ontology/1.

The reference architectures are the designs we publish. There are eighteen so far, across energy, oil and gas, ports, manufacturing, agriculture, rail and more, in the United States, Saudi Arabia and Pakistan. Together their object models hold 246 typed objects and 228 links. They are on codeatoms.ai and as a dataset on Hugging Face.

PADI, the Physical AI Design Index, is the benchmark. It asks a simple question: does the Praxis platform design better systems than frontier AI on its own? It scores both on fresh tasks across 13 industries in three countries. In our latest round, a blind evaluation on seven tasks the platform had never seen, the Praxis platform passed 48.5% of the checks against 28.4% for frontier AI, and came out ahead on all seven tasks. The results are on the PADI page.

Compression and Reinflation

The obvious way to give a model domain expertise is to fine tune it. Collect the data, train the weights, ship the model.

We chose not to, and I wrote about why in a memo, Compression and Reinflation. Here is the short version.

Fine tuning freezes what you know into weights. The day training ends, the world has already moved: a new regulation, a new site, a new country. And when a better model arrives, the money spent on the old weights is gone.

So we keep expertise as context instead. It works in two steps.

Compression. Every day, routines run over what the platform has learned: from real deals, from benchmark rounds, and from engineers correcting a design. Many observations become one sharp principle. Contradictions are resolved. What proved itself rises and what never mattered fades. What is left is small. It is not a record of everything that happened. It is the meaning of what happened.

Reinflation. When a new requirement arrives, say a copper mine in a country we have never worked in, no stored principle fits it exactly. It was never meant to. The model reads the compressed expertise and expands it into this site, this country's rules, this fleet and these people. The same lesson becomes a different design every time, because the world it lands in is different every time.

The idea comes from biology. The biologist Michael Levin describes memory as a bowtie: a lot of experience squeezed through a narrow middle and then expanded again for a new body and a new world. A caterpillar's lessons survive inside the butterfly, remapped for wings instead of legs.

There is one rule that makes this work. Everything is context for the model to reason over. Never a keyword match, never a lookup.

Why It Beats Fine Tuning

Keeping expertise as context instead of weights changes four things.

It stays current. A new regulation is one new entry, not a new training run.

It stays readable. An engineer can see what the system knows, where it came from, and correct it. Nobody can read a weight.

It survives the model. When a stronger open weights model arrives, the expertise carries straight over and the designs improve on day one.

It stays with the owner. The expertise lives inside the customer's own perimeter, on their own infrastructure. For a sovereign deployment, that is the whole point.

Back to the Truck

So what would it take to vibe code a truck?

An engineer would describe the operation: an autonomous haul fleet for a copper mine, in a named country, with a target output and a budget. The system would come back with the sensing, the compute, the models, the site network, the object model of the mine, the safety case outline, the approvals, the three year cost and the plan. Then the engineer would push back, as engineers do, and the system would explain itself and change.

That is not science fiction, and it is not finished either. It needs a model that is good at reasoning, which we now have. It needs deep domain expertise, carried in a form the model can read and an engineer can check, which is what we are building. And it needs a way to measure whether the designs are any good, which is why PADI exists.

When those three come together, the most complex physical systems in the world stop being the work of a handful of people with twenty years behind them. A young engineer anywhere can design them, and understand every choice they made.

Explore the designs, the benchmark and the platform at codeatoms.ai.

Physical AI System Design for Semiconductors: What the Published Designs Already Answer

Umar Bilal — Sun, 18 Oct 2026 02:54:19 +0000

First published on CodeNinja Atoms.

A fab asks the same first questions whether it is a front end wafer fab in the United States or a packaging and test plant anywhere else: what is happening across the site right now in one picture, which pump, chiller or scrubber in the subfab is drifting toward a stop, what is actually connected to the tool and facility networks and can that be proven to an auditor, does the level reading on a bulk chemical tank belong to the tank it claims, is the fire protection actually armed, who is standing near moving equipment, which shift is drifting toward an incident, and how many units left the building against what the records say shipped. The series on this site holds no semiconductor design yet. It holds eighteen open reference architectures in other sectors, and nine of them are written for an operation that asks one of those questions in another setting. This guide reads them as a fab operator would, using only what each paper says.

The map

The fab's question	The published design that answers it	Where it was written for
One picture of the operation from the systems it already runs	Port Twin	A landlord port authority in the United States
Which subfab or facility machine is about to fail	Turbine Island Watch	A combined cycle power plant in Pakistan
What is on the control networks, and the evidence an audit asks for	Grid Context Watch	An energy and utilities operator in the United States
Whether a bulk tank reading can be trusted	Loop Integrity Watch	A fuel terminal in Pakistan
Is the fire protection armed	Factory Fire Watch	High risk factories in Saudi Arabia
Who is near the equipment, on footage the fab owns	Pit Camera Watch	A vehicle proving ground in the United States
Which shift is drifting toward an incident	Sovereign HSE Watch	An oil and gas operator in Pakistan
How many units left the building	Steel Count Ledger	Steel mills in Pakistan, counted for a public authority
What the records already prove	Reliability Atlas	Desalination and treatment plants in Saudi Arabia

One picture from the systems already running

Port Twin is the design to read when the problem is that every dashboard is a slice. It binds 8 operational systems, 5 live sensor feeds and 1 finance backbone into 13 typed objects and 13 links with the berth as the focal object, joined through three adapter families and one event backbone. The rule a fab will recognise is the boundary: the twin shows, and pilots, berth planners, engineers and finance staff decide in their own systems, with nothing written back into a system of record. The only model in the register is one self hosted open model, BGE-M3, for document retrieval, about 1.1 GB of fp16 weights on the port's existing enterprise virtual machines, so no equipment is bought. For a fab, the manufacturing execution system, the facility monitoring system, the maintenance system and the document store take the places of the port's systems; the pattern of a read only twin over owned systems carries across unchanged, and the object at the centre is the one the fab chooses.

Subfab and facility equipment: advise, never trip

Turbine Island Watch reads a rehabilitated 280 MW hydrogen cooled steam turbine generator through a one way OPC UA read from the DCS, the vibration monitoring rack and the generator protection systems, eight named plant systems in all entering through four adapter families. Three of its rules transfer to a vacuum pump, a chiller, an abatement unit or a makeup air handler without change. The read is one way and nothing writes to the DCS. The platform advises and the protection layer trips, so every alert carries a prescribed action that a named shift operator acknowledges. And the rollout gate is a number: 2 false alerts per week, with thresholds tuned only on channels with provable history, because the paper names alert flooding as what trains operators to ignore a platform. MOMENT-1 (MIT) detects anomalies zero shot over the telemetry, IBM Granite 4.2 8B (Apache-2.0) drafts diagnostic notes and test reports, and no training run is needed. The compute is one 48 GB PCIe inference GPU class card on ruggedised edge servers inside the plant perimeter, costed at about 22,500 US dollars to own the site envelope over three years.

What is on the control networks, with the evidence

Grid Context Watch is the design for the question an OT security lead and a compliance lead ask together: what is connected to the control networks, whether anything on them is behaving abnormally, and whether that can be proven with evidence when the auditor arrives. It joins 10 source systems, among them the SIEM, an OT detection tool, an IT monitoring platform, the energy management system and the substation data platform, through 2 adapter families into 14 typed objects and 14 links with the OT asset as the focal object, including the vulnerability finding, the detection rule, the pcap evidence record and the CIP evidence artefact. The boundary is read only and one way out of the control networks; nothing writes into an electronic security perimeter, and no data leaves the operator. Detection stays in the procured sensors. GLM 5.2 (MIT) reasons and Qwen3-Embedding-0.6B (Apache-2.0) retrieves, on 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. Owning that node costs about 510,000 dollars over three years, about four fifths the deepest three-year AWS commitment. Every triage, case and risk acceptance carries a named OT security analyst; agents draft, the analyst decides. The paper's evidence artefacts are written for NERC CIP; a fab puts its own security standard and its customers' audit questions in that place, and the shape of tool and facility networks watched one way and proven by record carries across.

Bulk tank readings that fail loudly

Loop Integrity Watch answers a question that a bulk chemical or gas yard shares with a fuel terminal: is the level reading arriving at the central host continuous, undistorted and attributed to the right tank. The paper's own framing names the failure a fab's facilities team will know, a flatlined gauge reads as a plausible low level and a swapped gauge makes two tanks trade identities. The design covers 13 fuel tanks on 3 Modbus RTU loops, with 3 booster installations engineered from a measured signal survey inside explosion proof junction boxes, and a governed inventory ontology of 15 typed objects with the fuel tank as the focal object, joined from 2 named source systems through 1 integration adapter family. The model register is empty: every integrity check (stuck value, swap, distortion) is deterministic. The compute is two containers on the operator's own site application server, inside its OT boundary, and every data quality event is acknowledged and resolved by a named person, with the location engineer signing acceptance.

Fire and life safety, read only

Factory Fire Watch gives 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 panel contacts, PLC inputs and LoRaWAN into an IoT platform hosted in Saudi Arabia, with never a write path into certified life safety equipment. The object model holds 14 typed objects and 12 links with the monitored point as the focal object. The one model, IBM Granite Tiny Time Mixers (TTM-R2), about 0.85 million parameters, runs on CPU inside the platform, and threshold evaluation on the safety path is rule logic. No server or GPU is bought. The field equipment per factory is about 2,600 to 3,200 US dollars in list prices, and a monitoring officer acknowledges or escalates every safety critical alert. The paper's framing suits a fab well: a sensor that has stopped reporting still reads as normal, so the worst failure mode is silence that looks like safety. For a fab and its fire pumps and water reserves, the design is the same read only estate with the sites renamed.

People near equipment, on footage the fab owns

Pit Camera Watch is a fully on premises camera system for a vehicle proving ground: 40 to 60 poles of fixed thermal and visible camera heads on new underground single mode fiber, 80 to 120 streams, recorded on an N+1 recorder pair running Frigate NVR, each holding 30 days of footage on local disk inside an isolated network. Its human rule is the conservative one an EHS office may want first. Booth operators watch and decide, nothing raises an automated alarm, and only the safety office can export a clip. The one model, D-FINE under Apache-2.0 at about 8 to 124 MB of weights, searches the archive after the fact for a test vehicle, a support vehicle or a pedestrian, on one 48 GB L40S class card per recorder host in the conditioned server room and nothing at the poles. Two choices carry straight to a United States fab. Section 889 supply chain screens come before any purchase, and camera uptime is a first class acceptance metric. The paper costs the estate at about 474,000 to 871,000 US dollars to own over three years, mostly camera heads, with no per camera license at any pole count.

Incidents before they happen

Sovereign HSE Watch predicts health, safety and environment incidents by joining 8 systems: SAP EHS, PM and QM; SCADA; fire and gas; Vision AI cameras; SQL databases; scanned files; and the identity provider. Its 12 typed objects include the HSE incident, the near miss report, the corrective action, the sensor reading, the anomaly event and the agent recommendation, which is the shape an EHS team in a fab with toxic and pyrophoric gas systems already keeps on paper and in separate tools. Six self hosted open weight models carry the work, including GLM 5.3 for reasoning, Chronos-2 for forecasting, RF-DETR for vision, BGE-M3 for retrieval and PaddleOCR-VL 1.6 for scans, with GLM 5.3 held at FP8 on one node of eight 141 GB HBM class GPUs (753 GB of weights, 904 GB with headroom). Every recommendation is approved or rejected by a named person, and nothing executes on equipment. The paper is written for Pakistan, so its export licence gate does not apply to a fab in the United States; the sizing arithmetic and the human rule do.

Counting what leaves the building

Steel Count Ledger puts IP66 HDR cameras and one sealed GPU industrial PC at every installation point, counts product as it is made, and publishes the counts through a one way link into one record the authority owns. Counting survives a slow wide area link because it happens at the edge, tracker models keep one item from counting twice, and the weighbridge feed serves as acceptance ground truth. Three self hosted open models carry it: RF-DETR for detection, fine tuned per product type and mill, Roboflow trackers for identity across camera fields, and GLM 5.3 for the compliance work surface. Every discrepancy case is judged by a named revenue field officer; the system counts and reconciles and never assesses, and the mill never edits a count. The paper also states a break even: the cheapest closed model matches the owned node at about 33 users, and above that, ownership is cheaper. The same pattern counts trays, reels or boxes at a packaging and test plant's outbound dock against what the manufacturing execution system says shipped, with the detector fine tuned per product type and site as the paper describes.

What the records already prove

Reliability Atlas is the design to read before any sensor is bought. It joins each plant's work orders, trip history and design documents through 3 read only adapters, as exports rather than live interfaces, into a reliability object model of 13 typed objects the operator owns, among them the equipment, the failure mode, the downtime event, the criticality index and the spare part. Availability is computed by Monte Carlo simulation on reliability block diagrams built from each plant's own failure study, auditable to ISO 55000, with predictions accepted only by a named reviewer and root cause reports validated only by a named engineer. The model register is empty by design, no hardware is bought, and the study runs on the operator's own in country servers. A fab with a maintenance system and years of facility downtime records can run this study first and let it decide which of the designs above it needs second.

What every one of them keeps constant

Across all nine, four things do not vary. The system runs on hardware the operator controls, or in a cloud or region the operator chose. A named person owns every decision, and nothing executes on equipment. The object model is published as JSON in the hyper-ontology/1 format, so a fab can read the exact objects and links before building any of its own. And the cost is computed the same way in every appendix, from the paper's own sizing with every price cited, so the own or rent question has one answer per design rather than a slogan; the comparison across designs is in Own or Rent the GPUs.

What a first semiconductor design would have to name

None of the nine holds a wafer, a lot, a carrier, a process tool, a chamber, a recipe, a reticle, a die or a test bin as an object, and none reasons over how they join. No published design covers wafer defect detection or fab yield analytics yet: none links an inspection result to the tool, chamber and recipe that produced it, and none traces a yield loss back through the lot's history. A first semiconductor design adds that object model, and the join between tool, lot, recipe and yield with it. It is the design the co-build page for semiconductors is open for: the operator brings the operation, the systems that already hold its records and one engineer who knows the fab; the published designs above are the proof of what comes out.

Use the designs

Every object model named here is in the paper's package as ontology/objects.json, and the same rows are in the Vertical-Driven Architectures dataset and served live by the codeninja-research MCP server, so a coding agent can ask which published design holds a vulnerability finding or a data quality event object and read why. Each paper is citable by its DOI on its page. All were reasoned on CodeNinja Praxis and their object models import into Hyper Ontology.

Every design cited here is one row in the table of sovereign AI reference architectures, and every object model is browsable among the industrial ontology examples.

About CodeNinja

System design for physical AI: restricted crop detection from orbit, on one 48 GB node inside the Kingdom

Umar Bilal — Sat, 17 Oct 2026 02:39:05 +0000

This is the engineering summary of an open reference architecture. The paper, its object model as JSON and the model register are free to reuse under CC BY 4.0: https://codeatoms.ai/restricted-crop-monitoring-saudi-arabia/. The operator is described by class, never by name.

An agriculture and earth observation operator in Saudi Arabia enforces a ban on green fodder and limits cultivation to the licensed area, because groundwater is the scarce input. Checking every parcel by visit is impossible. Checking by imagery alone produces flags nobody can act on.

Here is how an 18 month service turns imagery into case files that hold up.

1. Join eight sources into fourteen objects

The national agricultural register, crop and well licences, service records and a mirrored Sentinel-2, Sentinel-1 and Landsat archive enter through one adapter family. Farm holding, field, licence, tasking order, image capture, detection flag, case file and field inspector are among the fourteen objects. A detection means something only once it is joined to the licence of the holding it sits on.

2. Two small open models

RF-DETR, fine tuned per region, finds the crop. Chronos-2, zero shot, forecasts a vegetation index baseline for each parcel, so a field that departs from its baseline stands out. Both are Apache 2.0 and together hold about 0.55 GB of weights.

3. One card, inside the Kingdom

Both run on one 48 GB L40S class inference node in the operator's own facility. Owning it costs about 15,200 dollars over three years, about two thirds of the deepest three year AWS commitment. Any rented option today sits outside the Kingdom and would move the imagery and the register across the border. The card class needs a US export licence, which the paper flags.

4. A person opens every case

Only the flagged minority reaches a named field inspector, who verifies on the ground before a case file opens. The model narrows the search. The inspector makes the finding.

Full design, figures, the cost appendix with every price cited, and the object model: the paper. DOI: https://doi.org/10.5281/zenodo.23186675

System design for physical AI: one flight of 4 band imagery and LiDAR, bound into a site ontology

Umar Bilal — Fri, 16 Oct 2026 02:45:13 +0000

This is the engineering summary of an open reference architecture. The paper, its object model as JSON and the model register are free to reuse under CC BY 4.0: https://codeatoms.ai/vegetation-mapping-lidar-us/. The operator is described by class, never by name.

An agriculture and earth observation operator in the United States needs a vegetation baseline for its site. Aerial surveys usually arrive as a folder of tiles and point clouds. Next season, someone rediscovers what was flown, at what accuracy, and under which contract.

Here is how one flight becomes a baseline the next season can compare against.

1. Fix the acquisition, then everything else

One manned flight within a week either side of 1 July. 3 inch orthoimagery in four bands, red, green, blue and near infrared, plus USGS Quality Level 1 LiDAR at a minimum of 8 pulses per square meter. The leaf on window and the pulse density are what make the vegetation analysis possible at all.

2. Make the flight mission the focal object

Sixteen typed objects, from the site and its units, watercourses and dams, through the camera, the LiDAR sensor and the GNSS and IMU chain, to the orthoimagery, the classified point cloud, the elevation models and the QA/QC accuracy report. The contract, the invoices and both project managers are objects too. Every product traces back to the mission, the sensor and the accuracy report behind it.

3. Stay inside the operator's GIS

Four named sources enter through two adapter families into the operator's ArcGIS Enterprise. Nothing runs in a cloud the operator does not control.

4. Serve no model

Vegetation analysis stays with the operator's own analysts. The model register is empty and no compute is bought. The survey firm prices the acquisition against the operator's own task table.

5. Accept against evidence

The operator's project manager accepts each deliverable against the QA/QC accuracy report, not against a delivery date. The next flight compares against a versioned baseline.

Full design, figures and the object model: the paper. DOI: https://doi.org/10.5281/zenodo.23186673

System design for physical AI: one agriculture data dashboard the operator fully owns

Umar Bilal — Thu, 15 Oct 2026 02:47:06 +0000

This is the engineering summary of an open reference architecture. The paper, its object model as JSON and the model register are free to reuse under CC BY 4.0: https://codeatoms.ai/farm-data-dashboard-pakistan/. The operator is described by class, never by name.

An agriculture and earth observation operator in Pakistan runs a federal productivity pilot. Its farm sensor feeds, its historical datasets and a big data and analytics repository sit in separate places, so nobody can drill from a season down to one field and one reading without a manual export.

Here is how that becomes one dashboard the operator owns outright.

1. Join three sources through two adapter families

A telemetry adapter reads the site sensor feeds. An integration adapter reads the analytics repository and the historical datasets. Both run as scheduled batch pulls. A streaming backbone was set aside, because the sensor cadences and volumes are not yet known and the prototype has to settle them first.

2. Twelve objects, one model

Site, field, crop cycle, sensor device, sensor reading, historical dataset, publication, dashboard user, access role, alert, report and data source connection. Which fields in this crop cycle raised alerts that nobody acknowledged is one traversal, not three exports.

3. Open source on the operator's own servers

The model is projected as a versioned schema in PostgreSQL with PostGIS. Nothing runs in a cloud the operator does not control. The source code and full intellectual property are handed over at the end.

4. No model is the right model here

The analytics are deterministic aggregations, drill down and reporting, reproducible from the operator's own data. The model register stays empty and no hardware is bought, because nothing in the requirement needs inference.

5. Prove the slope on one farm first

The rollout starts with a one farm integration prototype. It measures what each additional farm costs to integrate before the fixed budget is spent. Every alert is acknowledged by a named user under a role the operator assigns.

Full design, figures and the object model: the paper. DOI: https://doi.org/10.5281/zenodo.23186671

The Model Register Across Sixteen Designs

Umar Bilal — Wed, 14 Oct 2026 02:53:34 +0000

Originally published on the CodeNinja Atoms engineering blog.

A model register is the part of a system design that names every learned component the operation will run, with its licence, its size and the reason it was chosen over the alternatives. Sixteen open reference architectures on this site carry one, and every row of every register is in the public dataset as models.jsonl. This note lines them up.

Two things stand out before the table. Four of the sixteen designs have an empty register, and in each case the emptiness is a decision the paper defends rather than a gap. And no design holds a model under a licence with a field of use restriction, a revenue trigger or a copyleft clause that would reach the operator's own code. Weight ownership is the rule the whole corpus runs on.

The reasoning models

Five designs put a frontier class open weight model behind a work surface for the operation's own people and agents. All five chose a GLM model from the same family, self hosted at FP8 on one node of eight 141 GB HBM class GPUs, and the register says why in licence terms.

Model	Licence	Designs that chose it	The reason the register gives
GLM 5.3, 753B mixture of experts, FP8	Bespoke licence permitting internal commercial use with attribution and no revenue trigger	Sovereign HSE Watch, Steel Count Ledger, Terminal Pulse	Strongest open agentic model; purely internal use is exempt from the model as a service security review trigger; 904 GB of weights against one node of 1,128 GB leaves KV headroom
GLM 5.2, 753B, FP8, 1M context	MIT	Grid Context Watch, Feeder Firewatch	A plain MIT licence with no revenue trigger or security review clause; one node holds it

The split is not an accident. Where the operator wanted the strongest agentic behaviour, the design accepted a bespoke licence whose internal use carve out it had read. Where the operator was a utility or a cooperative with an auditor behind it, the design took the one generation older model under MIT and gave up nothing it needed.

The detectors and trackers

Among the designs that watch with cameras, five chose the same detector.

Model	Licence	Designs that chose it	The reason the register gives
RF-DETR, Nano to Large checkpoints, BF16, about 61 to 68 MB at 16 bit	Apache 2.0	Sovereign HSE Watch, Steel Count Ledger, Feeder Firewatch, Terminal Pulse, Fodder Watch	The practical sovereign answer to the AGPL gate; fine tuned on the operator's own footage; the operator owns and hardens the weights. Two designs exclude the XL and 2XL checkpoints because they sit under a platform licence whose field of use terms would compromise weight ownership
D-FINE, N to X tiers, BF16	Apache 2.0	Pit Camera Watch	An unconditional licence; the N through L tiers fit one 48 GB card; ahead of DEIM on matched compute
Roboflow trackers, ByteTrack and OC-SORT class	Apache 2.0	Sovereign HSE Watch, Structure Phase Watch, Steel Count Ledger	Stable identity across frames without reintroducing copyleft after an Apache detector; one billet crossing two camera fields counts once
Anomalib 2.7.0, FP32, exported to ONNX Runtime	Apache 2.0	Hangar Form Cell	Unsupervised fit on the facility's own good panels needs no labelled defect set

The pattern a developer should copy is the pairing: an Apache detector followed by an Apache tracker. The register in three designs says in so many words that the tracker was chosen so that the AGPL exposure of the better known tracking collections never enters the serving path.

The forecasters

Model	Licence	Designs that chose it	The reason the register gives
Chronos-2, about 120M parameters, about 0.48 GB at FP32	Apache 2.0	Sovereign HSE Watch, Terminal Pulse, Fodder Watch	Zero shot multivariate forecasting with covariates and no field of use restriction; thousands of independent per parcel series with no fine tuning dependency
Chronos-Bolt	Apache 2.0	Hangar Form Cell	Learns per arm, per stylus load baselines on the edge box and flags drifting pass signatures
IBM Granite Tiny Time Mixers TTM-R2, about 0.85M parameters, FP32, about 0.003 GB	Apache 2.0	Factory Fire Watch	A CPU only forecaster inside the operator's existing platform; the requirement asked for no server or GPU, so none was deployed

The Granite row is the one to notice. The design covers a hundred factories and runs its forecaster on the CPU the operator already has, because a model of under a million parameters was enough and the requirement forbade buying compute.

Retrieval and document parsing

Model	Licence	Designs that chose it	The reason the register gives
BGE-M3, 568M parameters, FP16	MIT	Sovereign HSE Watch, Port Twin, Structure Phase Watch	Dense plus sparse plus multi vector retrieval in one pass with an 8,192 token window; multilingual retrieval across a workforce tagged by language; in the port design it is the only model in the register
Qwen3-Embedding-0.6B, 32K window, about 1.2 GB at BF16	Apache 2.0	Grid Context Watch, Terminal Pulse	Whole playbooks, CIP procedures or a whole shift's gate record retrieve as single passages; runs beside the frontier model
PaddleOCR-VL 1.6, about 0.9B parameters, BF16	Apache 2.0	Sovereign HSE Watch, Corridor Guard	Strong on degraded multilingual scans; no user or revenue threshold; digitises a scanned package with human reconciliation before any value enters the record

The choice between the two embedders is a window question. BGE-M3 wins where the corpus is drawings, inspections and messages in several languages. Qwen3-Embedding-0.6B wins where a single document, a procedure or a shift log, must retrieve whole, and its 32K window makes that possible.

The four empty registers

Design	What the register says	Why
Reliability Atlas	None: the register is empty by design	No inference workload is contracted; the deliverable is the study, and an empty register recorded plainly outranks a speculative entry
Loop Integrity Watch	Zero entries; all integrity logic is deterministic thresholds and rules	The failure modes are physical and protocol level, so no learned model is needed, no weights are held and no licence question arises
Field Ledger	Zero models committed; a forecaster over sensor time series stays open work gated on the first milestone's evidence	The requirement commits aggregations, drill down and reporting, and no model was verified for this run
Baseline	None; the register is empty by decision	No agent or model layer was wanted; analysis stays with the operator's own staff in their own tools

A quarter of the designs needed no learned model at all. Each of those papers still carries the register, with its emptiness and the reason written down, because a reader deciding whether to copy the design needs to know that the absence was chosen.

The licence pattern

Across the twelve named models the licences are Apache 2.0 (nine models), MIT (two) and one bespoke licence read for its internal use carve out. Three things never appear: a field of use restriction on a model the operator fine tunes, a revenue or user threshold, and a copyleft licence in the serving path. Where a stronger checkpoint existed under a platform licence, the register excludes it by name and says why.

That is the whole method. Choose the model for the job, read the licence as carefully as the benchmark, and write the reason down where the next engineer will find it.

Use the register

The rows behind this note are the models.jsonl file of the Vertical-Driven Architectures dataset, one row per choice with the design, the choice, what was picked and why. The same file is served live by the codeninja-research MCP server, so a coding agent can ask which designs chose a given model and read the reason. Every design named above was reasoned on CodeNinja Praxis and its object model imports into Hyper Ontology.