AI / ML Stack
Data → Train → Deploy → Audit. Four layers, each constrained by the same three operating principles. We are precise about what ships and what does not.
Four layers.
- 01
Data
AVIX dataset (Korean airfield species + global migratory profiles) + STANAG 2103 reference scenarios. Doctrine-tagged at ingest so every downstream artefact carries provenance.
- AVIX dataset · 250K labelled frames
- CBRN scenario corpus · 40 STANAG-categorised cases
- Habitat profiles · LULC + species, per coverage cone
- 02
Train
YOLO-class detection + Re-ID heads trained on edge-friendly backbones (Jetson Orin Nano target). Decision-support layer trained on doctrine-grounded prompt corpora rather than synthetic augmentation.
- YOLO-class detector · 4 ms/frame target
- Re-ID tracker · multi-target persistence
- Doctrine prompt fine-tuning · TPE-001~006
- 03
Deploy
Edge inference at the runway / vertiport / hazard cell. No GPU clusters in the field — every model has to run on consumer-grade edge hardware to pass deployment.
- Jetson Orin Nano · 8 GB
- TensorRT-class quantisation
- OTA model updates · sandbox-gated
- 04
Audit
Every entity we publish carries a `_uamkt_extensions:doctrine_ref` tag. Allied operators replay our submissions against their own validation harness before promoting them to operational decision surfaces.
- Doctrine-ref payload on every entity
- Replay harness · 19/19 reference run
- Public sample payloads (Phase 2)
Three checks.
Each principle is a binary gate: pass or do not ship. The bar is high on purpose — partners who adopt our outputs inherit our gates.
- 01
Edge first
No model ships if it cannot run on Jetson-class hardware in the field. Cloud inference is a fallback, not a target.
- 02
Doctrine-tagged
Every output is annotated with the allied doctrine reference it was checked against. Untagged outputs do not reach the catalog.
- 03
Replayable
Every published entity is reproducible from its source frames + model version. Audits do not stall waiting for vendor support.