By Muhammad Umar. This is the engineering summary of an open reference architecture from CodeNinja Atoms, a physical AI system design for composite part inspection, built so the physical AI at the inspection bench runs on hardware the operator owns. The paper, its object model as JSON and the model register are free to reuse under CC BY 4.0: https://codeatoms.ai/composite-part-inspection-mro-saudi-arabia/. DOI 10.5281/zenodo.24048561. The operator is described by class, never by name.
The composite repair shop of an aircraft maintenance, repair and overhaul provider in Saudi Arabia repairs radomes, flight control surfaces, nacelle panels, fan cowls and cabin monuments for commercial single aisle and wide body fleets. When a fan cowl arrives with a disbond at a hinge fitting, the certifying engineer needs two answers at once: does the measured extent fall inside the allowable limit for that part number at that location, and what happened the previous times this part type showed the same damage in the same place. The evidence exists, but apart: photographs stapled to a paper traveller, scan files stranded on the scanner's own computer, and a findings log joined to nothing. So the same disbond on the same panel type is assessed from scratch every time, and turnaround stretches at the most skilled step in the shop.
1. Read everything, write nothing below
The design is one pattern end to end: systems of record below, one governed object model in the middle, applications and agents above, all inside the shop network. The records below are already certified, licensed and audited, so nothing writes back to any of them. The single write path is a named person's confirmation or signature, and it lands in the object model, never in a system of record. Customer records, scan images and manual content never leave the shop network, with no data center or colocation entertained, so the reasoning runs on GLM 5.3 self hosted inside the shop instead of a cloud model API. Nothing crosses outward; updates cross inward only, through Lidi over a hardware data diode into an on site Harbor registry.
2. Two adapter families, seven sources
Seven sources feed the model: the camera stream from the bench and the scan station, and six read only systems of record: the maintenance information system, the scan equipment's ultrasonic and thermography image files, each customer's technical records with per customer isolation, the structural repair manuals under licence, the material and shelf life stock system, and the quality findings log. Two adapter families are the sole path in: an integration adapter against the maintenance information system's live API, and a file and document adapter lifting scanner exports and parseable manual content. Bespoke read connectors were picked over an integration product, because the maintenance information system exposes its own API and the scanner exports files.
The adapter tier guarantees four things. No source system is modified. Provenance is recorded on arrival: every work order, scan file and manual clause carries the system, the timestamp and the extraction method it came from, and where a scanner's export lacks the calibration reference the evidence chain needs, the gap is recorded and shown rather than hidden. Per customer isolation is applied as data lands. Ingestion is idempotent.
Apache Kafka in KRaft mode carries every arrival to the object model, ordered by work order so a damage confirmation never overtakes the inspection that precedes it, delivered at least once with idempotent consumers that drop repeats, and retained until every consumer acknowledges. Chrony with an OCP Time Card grandmaster aligns the clocks, Frigate NVR records the three bench and scan station cameras, Keycloak holds identity on site, and Prometheus 3.13 LTS with Grafana watches the running system.
Allowable limits are extracted from the licensed manuals as objects only once the licence is confirmed in writing to permit it; the stated fallback is citing the manual revision in each draft.
3. Sixteen objects, one walk
The object model holds sixteen objects with twelve typed links on Hyper Ontology v0.9: customer, aircraft, part number, serialized part, work order, task card, inspection, scan, damage, damage class, allowable limit, repair scheme, material lot, technician qualification, finding and release. Damage sits in the focal position, and the serialized part carries the shop's own states: Received, In assessment, Awaiting scheme approval, In repair, Final inspection, Released or Scrapped. The question the shop cannot ask today, how many fan cowls of one part number showed a disbond at the same hinge fitting in the last year, is four links: part number to its serialized instances, to the damage confirmed on them, to the scans and hinge fitting locations that evidence each one, and out to the findings log. A document store holds each of those facts somewhere. It holds no path between them.
4. Three tiers, sized by arithmetic
At the edge, two fanless sealed units of the octa core 6 TOPS NPU class sit at the inspection bench and the scan station, running ONNX Runtime. OWLv2 detects the shop's own damage vocabulary zero shot with no labelled local data, on selected frames; SAM 2.1 turns its boxes or an operator click into the masks and extents a damage map needs; Anomalib flags where an ultrasonic or thermography scan deviates from reference panels. The weights are small, so capacity is not the open question at this tier; rate is, and a dev board run with the actual models is the phase one gate. A whole shop camera rollout was set aside, because the value concentrates where damage is mapped and the NPU loads are unmeasured until that gate.
The site tier is one 48 GB PCIe inference card on a shop server running vLLM. It holds Qwen3.8-27B dense multimodal at about 28 GB of weights at 8 bit, reading scans and bench photographs into structured damage, and BGE-M3 at about 1.1 GB for retrieval over manuals, records and confirmed maps. Reading images here keeps them beside their data instead of sending them to the frontier node.
The central tier is one node of eight 141 GB HBM GPUs. GLM 5.3's 753 billion filed parameters at FP8 are 753 GB of weights; times 1.2 for the KV cache and activations that is 904 GB, and the node offers 1,128 GB. The node and the card are US origin accelerators for a D:3 and D:4 destination, so an export licence determination is a phase one checkpoint before any GPU is quoted, with sourcing inside Saudi Arabia weighed as the alternative. The latency budget is the shop's own: the frontier draft is waiting in the confirmation queue before the certifying engineer opens it. If an edge link fails, the units keep detecting and drafts queue locally.
5. Agents draft, engineers sign
Five services (one model of the shop, physical AI at the bench, named agents and memory, command center, assurance and acceptance) carry the work onto eight surfaces, among them the confirmation, scheme approval and release queues. Three agents draft on GLM 5.3: damage mapping, repair scheme and release paperwork. The certifying engineer confirms, amends or rejects the damage map and its classification, the design engineer approves or supersedes the repair scheme, material kit and qualification, and certifying staff sign the release. No agent classifies damage as repairable or releases a part. The disagreement path is built first, and the count of engineer reports is published weekly. Every confirmed map and approved scheme is kept as memory the next assessment of the same part type reads; nothing is trained into weights, and every model carries terms that let the operator hold it: the bespoke GLM-5.3 licence, Apache-2.0 or MIT for the rest.
Appendix A prices all three tiers at about 613,000 US dollars to own over three years (583,000 to 639,000: about 570,000 for the node, about 42,000 for the site server and card, under 1,000 for the two edge units, with support and power at the commercial electricity tariff). Renting the node and the card on the deepest three year commitment in AWS Middle East (UAE) is about 773,000, so ownership costs about four fifths of it; on demand it is about 2.06 million, against which ownership costs about three tenths. Renting is a price reference only: AWS listed no Saudi region on the day the prices were read, and every rented option would carry customer records, scans and manual content across the border.
The rollout is two phases and 18 scope items (17, then 1) across five workstreams, closed by gates, never dates, and traced against 22 requirements. Phase one runs on fan cowls and nacelle panels for one quarter: baseline first, then shadow mode with no draft holding a required record, then a capability demonstration on reference panels, with the quality system change under the civil aviation regulations and the personal data protection assessment filed in parallel. Its exit gate is the dev board run of the actual vision models on the 6 TOPS NPU class. The measures are assessment hours per part, turnaround days, repeat findings on the same part type and rework after final inspection. Phase two extends to flight control surfaces and the remaining part families.
Made with 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/composite-part-inspection-mro-saudi-arabia/
One of the CodeNinja Atoms worked examples of physical AI system design: systems of record read and never written, a governed ontology joined by typed links, open weights on hardware the operator owns in its own country, and agents that draft while named people approve, the shape every physical AI design in the series is published in, with its object model as hyper-ontology/1 JSON under CC BY 4.0.
Top comments (0)