The image
A vendor BSP you can't rebuild
The SDK image that made the demo work was flashed once. Nothing records how it was built, so nobody can reproduce it.
EDGE AI INFRASTRUCTURE
Bring AI applications to production hardware faster
Production-ready Linux, supported accelerators, prebuilt AI packages, and repeatable deployment workflows, so AI workloads move from development into the field.
The gap
Prototypes don't survive the move to production
The hard part of edge AI is not getting a model to run once. It is keeping that workload buildable, shippable, and updatable on hardware you do not control.
Prototype environment
Vendor SDKs and desktop distros get a demo running fast, then stop being the thing you can ship.
The AI stack
Libraries pinned by hand
Runtimes, drivers, and Python packages installed by hand on whichever machine was nearby. Every box drifts from the last.
The lifecycle
No path after the demo
No way to get a better model onto a device already in service, and OS maintenance and CVE response land on your team.
Production reality
Avocado gives the same workload a maintained base you can build, ship, and update.
The image
One declarative build
The same configuration produces the same image on every machine, including the one that builds your release.
One imageThe AI stack
The stack is part of the image
Declare the accelerator runtime and the AI libraries beside the rest of the system. The build resolves them once and records the result.
DependenciesThe lifecycle
Updates are a service
Ship models and applications over the air from Connect, on a base that someone else patches.
In the fieldSupported hardware
From a 0.5 TOPS microNPU to a 275 TOPS Jetson
Pick the silicon your product needs. Every target below builds from the same declarative configuration and the same CLI. There is no per-vendor toolchain to stand up first.
NVIDIA Jetson
NXP i.MX
Renesas RZ
Synaptics
NVIDIA
Orin Nano
67 TOPS · Ampere GPU, 1024 CUDA cores
NVIDIA
Orin NX
100 TOPS · Ampere GPU, 1024 CUDA cores
NVIDIA
Jetson AGX
275 TOPS · Ampere GPU, 2048 CUDA cores
NXP
i.MX 8M Plus
2.3 TOPS · Integrated NPU
NXP
i.MX 93
0.5 TOPS · Arm Ethos-U65 microNPU
NXP
i.MX 95
eIQ Neutron NPU
Renesas
RZ/V2H
Integrated AI accelerator
Synaptics
Grinn AstraSOM-1680
Integrated NPU
No board yet? Build against QEMU (virtual), then change one line to a real target.
View all supported hardwareDependencies
Use the libraries and packages your workload needs
Declare the AI stack alongside the rest of the system. The build resolves it once, records it in the image, and produces the same result on every machine.
Declare once
Put the AI stack and the rest of the workload in one manifest.
Resolve at build
The build pulls from pinned feeds and records what landed in the image.
Same everywhere
Every machine gets the same resolved result, no device-side installs.
One image
Build on the same image you ship
Local iteration and the production artifact are the same declarative build. The AI stack your application was developed against is the one that boots in the field.
One declarative build
Declared in one config, not assembled by hand on each machine. How Avocado OS builds the image
Reproducible months later
Packages resolve from pinned feeds, so the same config rebuilds the same image.
Bench hardware is ship hardware
Flash the signed image to your dev kit, then to the units you ship.
01
Declare the workload
Name the accelerator, the AI packages, and your application. Nothing about the stack is left to a machine’s history.
Local to hardware
$ avocado install
avocado build
avocado provision02
Build it, then run it on hardware
Three commands take the declared config to a signed image, then onto the dev kit in front of you.
03
Update it in the field
Models and applications move over the air, and only the layer that changed gets pushed. See how Connect ships updates
In the field
Deploy and operate AI workloads after launch
A model that improves is only useful if you can get it onto hardware already in service. Peridio Connect handles the rollout, the segmentation, and the access.
01
Model and application updates
Ship a new model or a new application version without rebuilding and reflashing the whole system.
avocadoconnect uploaddone
Composed image
Application + modelchanged
CUDA · TensorRTunchanged · deduplicated
Avocado OS baseunchanged · deduplicated
100%
Device partitions · A/B
Slot A
v1.2.2
standby
Slot B
v1.2.3
active
02
Fleet segmentation
Roll out to a cohort first (canary, staging, then production) instead of every device at once.
avocadoconnect deploydone
canarypromoted
stagingpromoted
production100 devices
03
Rollout visibility
Watch an update land device by device, and stop or roll back when a cohort reports trouble.
rollout · v1.2.3done
v1.2.3production
Rollout complete- dev-01
Updated
- dev-02
Updated
- dev-03
Rolled back
- dev-04
Updated
04
Remote diagnostics
Open a tunnel to a specific device to inspect a failing inference run on the hardware it happened on.
avocadoconnect tunneldone
Secure session
youedge-cam-04
- Device
- edge-cam-04
- Target
- jetson-orin-nano
- Tunnel
- wireguard · encrypted
- Session
- logged
- Fault
- /dev/video0 busy
Workloads
What teams are putting on this hardware
Different workloads, the same infrastructure underneath. Follow a card for the depth on that one.
Computer vision
Camera pipelines running detection or classification on-device, where inference has to keep up with the sensor rather than a round trip to a server.
Jetson Orin · i.MX 8M Plus
Read more
Robotics perception
Perception and navigation stacks that need GPU-accelerated runtimes and a maintained base OS underneath them.
Jetson Orin NX · Jetson AGX Orin
Read more
Industrial inspection
Fixed-position quality and defect detection on the line, where a device may run for years between physical visits.
i.MX 95 · Jetson Orin Nano
Read moreNext step
Choose a starting point
Four ways in, depending on how far along you already are.
Check the hardware
Is my target supported?
The support matrix lists every validated board, the accelerator on it, and what ships pre-validated.
View all supported hardwareStart building
I want to try it now
Go from a clean machine to a signed image running on a dev kit, following the documented path.
QuickstartBuilding a robot
My workload is robotics
Perception stacks, Jetson specifics, and the lab-to-field problem get their own treatment.
Go to RoboticsTalk it through
I have constraints to check
Bring your accelerator, your runtime, and your timeline, and we will tell you what is validated.
Talk to an engineer