Lamantin AI
Infrastructure

fluxion

A cross-hardware SDK for making sense of media streams.

Rust·not yet public

Not open yet — the substrate ships as it hardens

One pipeline graph, swappable hardware underneath.

Not a monolithic runtime: a set of GStreamer plugins on an SDK foundation, from which real pipelines get assembled. The idea it is built around is that an application's pipeline graph should stay stable while the runtimes and the silicon underneath it change — one pipeline contract with replaceable intelligent elements, rather than a universal model server.

Three layers. flxcore carries metadata, logging, inference contracts and the backend libraries; flxgst exposes those capabilities as plugins inside a streaming graph; flxpython binds the metadata for Python around native classes rather than raw GObject wrappers.

It is deliberately not an orchestrator for every AI workflow, and not a collection of speculative backends before the pipeline primitives are finished.

  • Inference Detection, pose estimation and instance segmentation on YOLOv8, over ONNX Runtime on CPU or Rockchip RKNN on an RK3588 NPU — with the backend chosen at compile time.
  • Tracking Online tracking with stable per-object ids, on jamtrack-rs and ByteTrack.
  • Metadata that survives the graph Primitives for frames and detections that propagate across muxing, inference, tracking and rendering instead of being rebuilt at each step.
  • Measured, not projected Four parallel YOLOv8 streams with tracking and overlay at 18 fps on an RK3588.

Eyes that already think.

Fluxion is meant to be one of the sensory sources feeding octo: its connector emits high-level scene events rather than frames, and octo's reflex layer answers them deterministically or escalates to cognition.

the graph 
camera → flxgst pipeline
    ├ flxneuro   // detection · pose · segment
    ├ flxtracker // stable ids
    └ flxdraw    // overlay
         ↓
    scene events → octo