THE DEPLERO PLATFORM

One platform for every model,
at every site.

Deplero deploys, updates and watches your models across customer sites, on-prem GPU clusters and cloud accounts on any provider. It's installed in your environment, connected to your devices and models, and run by your team.

What it does

See what runs where. Ship every release safely.

Fig. 1. The console overview, shown with sample data: every site with its devices and model version on the left, the GPU clusters behind them on the right, and a release rolling out to the sites.

Every model, every site

Which version runs on which device, at every customer site, in one view. Drift shows up before a customer notices it.

Signed releases

Each release is signed in your registry and verified on the device before it installs.

Capacity check

A new model only goes where the GPU has room for it. Devices that don't fit stay on the current version.

Shadow test on live data

The new model runs beside the current one on each site's own data, and goes live only when it clears the bar your team sets.

Staged rollout, automatic rollback

One device first, then wave by wave. If a release fails anywhere, every device goes back on its own.

Edge to cloud, one system

Edge devices, on-prem GPU clusters and your AWS, Azure or GCP accounts, with data rules set per site. Sites keep running when the network drops.

The release path

Every release takes the same path to every device.

Signed, verified, checked for GPU room and shadow-tested on live data, then rolled out one device, one site and every site. If any device fails its shadow test, every device returns to the current version on its own.

  1. SignedRelease signed in your registry
  2. VerifiedEach device checks the signature before install
  3. Capacity checkGoes only where the GPU has room
  4. Shadow testRuns beside the current model on live data
  5. Wave 1One device
  6. Wave 2One site
  7. Wave 3Every site

Release live on every site. One device held on the current version: not enough GPU room.

Fig. 2. Every release takes the same seven steps to every device: signed, verified, checked for GPU room, shadow-tested on live data, then one device, one site, every site. A device without room is held on the current version; a failed shadow test sends every device back on its own.
Fourteen sites during a rollout, chapter: Roll backFourteen sites drawn as dots in three rows, Site 01 to Site 14. Pale dots run the current version, teal dots run the new release, an amber dot is held and reverted, and a green ring marks a site that passed the gate. Sign: all pale. Stage: Site 01 teal. Gate: Site 01 teal with a green ring. Roll out: Site 01 to Site 07 teal. Roll back: Site 01 to Site 06 teal, Site 07 amber, Site 08 to Site 14 pale.Site 01Site 02Site 03Site 04Site 05Site 06Site 07Site 08Site 09Site 10Site 11Site 12Site 13Site 14
  • current version
  • v2.4.1
  • gate passed
  • held

activity: Site 07 held, reverted; 6 sites stay on v2.4.1

One site fails its check and goes back to v2.4.0 on its own. The rest keep the release.

Fig. 3. The same rollout in five chapters your team can step through: sign, stage on one site, pass the gate, roll out wave by wave, and roll back. When a site fails its check it goes back on its own and the rest keep the release.

Release status

v2.4.1published, verified, rolling:9 of 14 sites,1 held,0 failed

Fig. 4. The line your team watches during a rollout: the release is published and verified, nine of fourteen sites are on it, one is held and none have failed. Sample data.

How it works

Three parts, all in your environment.

Fig. 5. From your side there are three things: what you ship, one release signed by you; where it runs, identical sites inside your environment; and what you see, one console with every site on the same version and the one held site visible.

AT EACH SITE

The Deplero agent

A small agent beside your model server on every edge device. It pulls signed releases, verifies them, runs the capacity check and the shadow test, and keeps serving if the network drops.

IN YOUR CLOUD OR DATACENTER

The console and release engine

The console, your model registry and the rollout engine, running in an account or datacenter you control. This is where your team sees every site and ships every release.

YOUR GPU CLUSTERS

Heavy models and retraining

The larger models and retraining jobs, on-prem or spread across your cloud accounts, fed by each site under the data rules you set for it.

DEPLOYMENT 01

Edge + cloud

Vision at the edge. Heavy compute where it belongs.

For products whose inputs come from the physical world (cameras, scanners, robots, sensors). Real-time inference runs on the device next to the data; the heavier models run in your datacenter or cloud. The platform connects the two and keeps every site on the right release.

WHAT'S IN IT

  • Edge inference on Jetson-class devices, edge servers and on-prem GPU boxes
  • Serving for custom, vision and multi-modal models
  • The split between what runs at the edge and what runs in the core
  • Staged rollout and rollback across every customer site

DEPLOYMENT 02

On-prem & air-gapped

The full platform in a datacenter, even with no internet.

The Deplero platform inside a datacenter or colo, from bare metal up. For data that can't leave the building, latency that can't take a round trip, or GPUs that are bought and waiting for production.

WHAT'S IN IT

  • Bare metal to production Kubernetes, drivers and networking included
  • GPU scheduling, storage, model serving and monitoring
  • Fully air-gapped installs, with releases mirrored inside the network
  • Everything as infrastructure-as-code in your repositories

DEPLOYMENT 03

Multi-tenant

Shared GPU capacity, every customer kept separate.

For products that run other people's workloads: AI vendors hosting customer data, and teams running shared GPU capacity for many groups. The hardest layer in the stack, built into the platform.

WHAT'S IN IT

  • Tenant isolation across compute, network and data
  • VM workloads alongside containers
  • Node health you can trust, not just node health that reports green
  • Scheduling, quotas and capacity planning across clusters

The stack

Boring, proven, open.

Built on the open tools your team already runs. What Deplero adds is the agent, the release path and the console on top of them. Configuration lives in your repositories as infrastructure-as-code.

Bare-metal KubernetesNVIDIA TritonvLLMJetson / edgeKubeVirtMIG / fractional GPUMulti-tenant isolationTerraform + GitOpsPrometheus + GrafanaAir-gap delivery

Where it runs

Infrastructure you control. That's the only requirement.

Server rack in a datacenterA rack elevation drawn in outline with four installed GPU servers, each with status lights, and two empty slots: the customer's own datacenter or colo.

Your datacenter or colo

From bare metal to production: OS, drivers, Kubernetes, GPU scheduling, storage, serving, monitoring. Fully air-gapped installs supported.

Edge inference deviceA small fanless edge box with a breathing status light, drawn in outline: inference on the device, at the site.

The edge

Inference at the edge, where your data is born: edge boxes and on-prem GPUs next to the cameras, robots, scanners, and sensors. Heavier work is handed back to your datacenter or cloud.

GPU instances inside your own cloud accountThe same GPU server drawn inside a dashed boundary labelled your account: the hardware is the cloud provider's, the account around it is yours.YOUR ACCOUNT

Your cloud tenancy

GPU workloads inside your own AWS, Azure, GCP or OCI accounts, via infrastructure-as-code you keep. Several providers at once if that is where your credits or your customers are. You pay each provider directly.

Runs on
  • Jetson Orin
  • x86 RTX
  • L40S
  • air-gapped
  • offline for weeks
Serves
  • Triton
  • ONNX
  • TensorRT
  • vLLM
  • containers
  • Kubernetes
Where
  • hospital
  • plant
  • robot cell
  • warehouse
  • lab
Fig. 6. Where the platform runs, what it serves and the settings it is used in. The same release goes to every one of these the same way.
Jetson Orin NX module seated on a green carrier board, with the system-on-module and its connectors in sharp focus
Sample status overlay
  • edge device, enrolled
  • model v2.4.1, verified
  • serving, offline tolerant
Fig. 7. A Jetson Orin NX module on a carrier board, the class of edge device the platform runs on next to the cameras, with a sample status overlay. Photo: 4300streetcar, Wikimedia Commons, CC BY 4.0, cropped
Open x86 workstation with four RTX 6000 Ada cards installed beside the CPU cooler and memory
Sample status overlay
  • GPU box, enrolled
  • 6 devices, one view
  • no outbound to vendor
Fig. 8. An x86 workstation with four RTX 6000 Ada cards, the class of on-prem GPU box that serves the heavier models at a site, with a sample status overlay. Still: Geekerwan via Wikimedia Commons, CC BY 3.0, cropped

Getting started

Scoped and priced before anything starts.

01

Assess

Short · fixed fee

We review your sites, devices, models, data rules and hardware, owned or planned. You get an architecture, GPU sizing and a plan for bringing the platform up, with a price.

02

Install

Fixed scope · fixed price

We install the Deplero platform in your environment, connect your devices, models and GPU compute, and prove it against acceptance criteria agreed upfront: latency, throughput, failover.

03

Run and grow

Yearly platform license

Your team runs it, with runbooks and training. The license covers new platform releases and direct engineering support. Every new site and every new model version goes through the same platform.

FAQ

Straight answers.

Are you a software company or a consulting firm?

A software company. Deplero is a platform. We install it in your environment and train your team, and a yearly license covers new releases and support.

What runs on the device?

A small Deplero agent next to your model server. It pulls signed releases, verifies them before installing, runs the capacity check and the shadow test, and reports health to the console in your environment. It makes no connection to Deplero.

How do updates reach an air-gapped site?

As signed release bundles that go through your own controlled transfer process into a registry mirror inside the site. Devices verify the signature before they install anything.

If nothing phones home, how is the license tracked?

The license is a yearly agreement based on the number of sites. The platform doesn't call Deplero to check it, and it keeps running either way.

What happens if we stop working with Deplero?

Your platform keeps running. It has no license server and no dependency on Deplero's systems, and its configuration lives in your repositories as infrastructure-as-code.

Does it replace our device fleet management tool?

No. Deplero handles models and releases. Device inventory, OS patching and hardware health stay with the fleet tool you already use.

Where does the platform run?

In your environment: at your customer sites, in your datacenter or colo, and in your cloud accounts. Your data stays there, and you pay your cloud provider directly.

Can we use more than one cloud?

Yes. Each cloud account is a site on the same platform, so a tenant can span AWS, Azure, GCP and OCI at the same time. Startups use it to spend credits from two providers on the same models. Vendors use it to land a customer on the cloud that customer mandates, without a second release path.

Who runs it day to day?

Your team. The platform is built so your engineers can run it, with runbooks and training from us. The platform license includes new releases and direct engineering support.

We don't own GPUs yet. Too early?

No. Earlier is cheaper. The assessment covers hardware and device selection and sizing before you buy, which often saves more than the assessment costs. We're vendor-neutral on hardware: we don't resell it and take no commissions.

Which models do you deploy?

Your own fine-tuned, vision, or domain models, plus any open-weight model that fits your hardware and use case. Deplero runs and ships them; we don't build or clinically validate the models themselves.

Will you build our AI application, agents, or models?

No. That's a deliberate line. Deplero is the platform your developers (or your application partner) ship on. It keeps us focused, and it means we never compete with the teams we support.

Start with an assessment.

Two weeks, fixed fee, and you keep the architecture and sizing plan whether or not we build it. The cheapest mistake-prevention in the project.

Book a demo