AI Funnel — describe your edge AI flow in TOML.

Build phase of the AI funnel deployment lifecycle, inside Cloud Connect. The customer container — funnel.toml, models, COCO dataset, and container code — is pushed to a registry that is either Munic-hosted or customer-hosted. From the registry, the artefacts feed the triage trainer, which merges multiple customer funnels into one unified triage detector. The result feeds the NPU optimizer, which enforces NPU-only operations and low-precision inference. The optimized model feeds the hardware-in-the-loop validator, which runs the candidate against reference target hardware and gates on accuracy and latency. A bundle that passes is signed and emitted as the deployment bundle.

flowchart LR
  CB["Customer container<br/>funnel.toml + models<br/>+ COCO dataset + container code"]
  REG[("Registry<br/>Munic-hosted or<br/>customer-hosted")]
  TT["Triage trainer<br/>multiple funnels → one triage"]
  OPT["NPU optimizer<br/>NPU-only ops · low-precision inference"]
  HIL["HW-in-the-loop validator<br/>reference target hardware<br/>accuracy + latency gate"]
  BUNDLE["Signed deployment bundle"]
  CB --> REG --> TT --> OPT --> HIL --> BUNDLE
  class TT,OPT,HIL ai-node

Provision phase of the AI funnel deployment lifecycle, on the device. The signed bundle is pulled through the same OTA channel as code updates. The DAG compiled from funnel.toml is dispatched, and three configurations fan out in parallel: camera and sensor routing, the AI runtime loading the models into the NPU, and the GPU ROI shader configuring tensor format, letterboxing, and ROI extraction. All three converge at a ready signal — the device is armed for the first frame.

flowchart LR
  PULL["Bundle pulled<br/>same OTA channel as code"]
  DAG["DAG dispatched<br/>compiled from funnel.toml"]
  CFG_CAM["Camera / sensor<br/>routing configured"]
  CFG_NPU["AI runtime<br/>models loaded into NPU"]
  CFG_GPU["GPU ROI shader<br/>tensor format · letterbox · ROI"]
  READY["Ready · first frame"]
  PULL --> DAG
  DAG --> CFG_CAM --> READY
  DAG --> CFG_NPU --> READY
  DAG --> CFG_GPU --> READY
  class CFG_NPU ai-node

Run phase of the AI funnel deployment lifecycle, per frame on the device. The camera or sensor produces a shared GPU frame. The GPU produces a triage tensor. The NPU triage model emits labels and bounding boxes to the AI funnel DAG router. The router dispatches per detection: the GPU re-crops the region of interest with letterboxing and normalization, the NPU runs the downstream model and returns labels and confidence. The router either loops back to chain another model in the DAG or exits to the customer container, where the workload terminates as tracking, mapping, a cloud message, or a CAN message — the terminal node depends on the DAG.

flowchart LR
  SRC["Camera / sensor<br/>shared GPU frame"]
  GPU1["GPU<br/>triage tensor"]
  NPU1["NPU triage<br/>labels + bboxes"]
  AIF["AI funnel<br/>DAG router"]
  GPU2["GPU<br/>per-detection ROI<br/>letterbox · normalize"]
  NPU2["NPU model<br/>labels + confidence"]
  SINK["Customer container<br/>tracking · map · cloud message · CAN"]
  SRC --> GPU1 --> NPU1 --> AIF
  AIF --> GPU2 --> NPU2 --> AIF
  AIF --> SINK
  class NPU1,NPU2 ai-node