Parallax
Platform

One platform, from the sensor to the decision.

Parallax is four parts that are designed together: models, a runtime, a desktop app and an SDK. Each one assumes the others are there, which is how a whole perception graph fits inside 8 ms.

The parts

Four pieces, one frame budget.

Use them together, or bring your own model and use the Runtime and SDK on their own.

Parallax Runtime

The engine. Compiles a perception graph for one chip and runs it one frame at a time, every frame.

How the Runtime works →

Depth models

Detection, masks, tracking, pose, stereo depth and event motion, sized to fit on-chip memory.

See the models →

Parallax Studio

A desktop app to add cameras, draw zones, watch detections live and read traces frame by frame.

Inside Studio →

SDK

Swift, Kotlin, C++ and Python. The same session, camera and event types in every language.

Read the docs →
Parallax Runtime

A runtime that treats every frame as a deadline.

Inference servers wait for a batch. The Runtime does not wait for anything. It compiles one schedule per chip and keeps the answer on time, even when a node runs long.

Capture
Frames come off the sensor and through the image processor without a copy.
Schedule
Each node gets a slice of the frame budget, measured on this chip.
Infer
Fused int8 kernels run from on-chip memory, on the NPU, GPU or DSP.
Decide
Policies turn detections into events before the next frame arrives.
What the Runtime guarantees

Late answers are worse than no answer.

So the Runtime never delivers one. If a node overruns, its last good result is handed on and marked stale.

On time

Every frame gets a decision inside its budget. Overruns are handled by policy, never by a queue.

7.8 ms median · 9.6 ms p99 on dock-03

Per chip

Schedules are compiled for the exact chip, so fusion and memory layout fit the hardware, not a framework.

5 chip families

Offline

No licence check, no telemetry, no model download at run time. It works in airplane mode.

0 frames leave the device
Parallax Studio

See what the device sees, while it sees it.

Studio connects to a camera over the local network or USB and shows the graph running on the device itself.

Parallax StudioOn-devicedock-03 · Depth-S · int8
Loading dock camera feed
128 fps7.8 ms median4 tracksairplane mode
  • Live feed with tracked boxes, labels and per-frame timing.
  • Zones and rules drawn on the image, pushed to the device in one click.
  • Traces that show where each frame spent its budget.
  • Works fully offline. Nothing in Studio needs a network connection to the camera.
Where it runs

Built for machines that move, and cameras that watch.

The same Runtime and models ship on every device class. Only the compiled schedule changes.

Fixed cameras
Loading docks, gates, aisles and production lines. Events go to your own systems, video stays on the camera.
Depth-S · Depth-Track
Mobile robots
Obstacle and person detection with metric depth, inside the control loop instead of beside it.
Depth-Stereo · Depth-S
Glasses and wearables
Hand and body keypoints at display refresh rate, on a battery budget.
Depth-Pose · Depth-Event
Kiosks and checkout
Item detection with instance masks and on-screen feedback that never waits for a server.
Depth-M
Privacy by architecture

The video never has to leave the building.

Parallax sends events, not pixels. What leaves the device is the decision, in a few hundred bytes.

  • Frames are processed in memory and discarded unless you choose to record.
  • Events carry boxes, tracks and labels, not images.
  • Models and schedules are signed and verified on the device.
  • Runs on isolated networks with no outbound connection at all.

Put real-time vision on your device.

Install the SDK, point it at a camera and see your first detections in an afternoon.