The primitives
Functions for short work, sandboxes for untrusted code, agents for the conversation.
AI for partner networks
AI for the people who sell for you
Each partner gets their own agent — memory, tools, and a spend cap. Their data does not mix. You keep the brand.
Product
Four parts. One system for the channel.
An agent per partner, with its own tools and memory. Their customers never see another partner’s context.
The models you already meant to use
Gemini, Gemma, or a compatible endpoint. No rewrite of the partner’s stack.
Shaped to the traffic
Chat, batch, or a spike at month-end. The endpoint follows the load.
You keep the face of it
We run the runtime. The partner sees your product, not ours.
Fine-tune when you have the data. The job runs, you get a versioned endpoint back.
Ordinary training jobs
SFT or LoRA. You send the dataset. We return a version you can promote.
No cluster to babysit
Jobs go up on Cloud Run or GKE and come down when they finish.
One GPU or a few
Ask for the size the job needs. Don’t pay for a rack you are not using.
A short-lived machine for an agent that needs to run code or open a file. Snapshot it. Throw it away.
Many at once
Spin up a lot of sandboxes when a rollout needs them. They do not share disks.
Your image or ours
Start from a default image or the one the channel already uses.
It shuts itself off
CPU and memory follow the job. Idle time does not sit on the bill.
A function becomes an HTTP endpoint. It grows with traffic and goes quiet when nobody calls.
Work in parallel
Fan out a batch without writing your own queue.
Long jobs in the back
Leave a run going. Checkpoint lands in Cloud Storage.
A door on the front
Auth, a cap per partner, and a log you can read later.
Platform
What you need to ship AI to a channel
See the productFiles next to the job
Weights, datasets, and images sit close to where the run happens.
Containers that start fast enough
A runtime for large images and a queue when a job needs an accelerator.
Where it runs
São Paulo first. Other regions if you need them.
Talk to us
São Paulo southamerica-east1
Iowa us-central1
Belgium europe-west1
Close to the channel
Default region is São Paulo. We can stand up another Google Cloud region if a customer asks.
Scale with the partners who are actually on
Idle tenants should not keep machines warm.
Training when you need it
Training jobs can take a GPU or TPU on Vertex AI when you need them.
You can see what broke
Logs and alerts per tenant, so a partner issue stays a partner issue.
In production
The boring parts that keep a channel safe
Access control, logs, and a bill you can show a partner.
10 Sep, 13:54sandbox.createpartner-bot
10 Sep, 13:32deploy.terminatedeploy-bot
10 Sep, 13:14image.pushci-runner
10 Sep, 13:02volume.deleteorby-bot
Agents0
Utilization0%
Tokens / min0k
One partner, one box
IAM, audit logs, and data that stays where you put it. Built with LGPD in mind.
Logs you can hand over
What ran, who called it, what it cost. Export if you already have a sink.
A cap, then the meter
Pay for what the code uses. Set a ceiling per partner so a runaway job does not surprise you.
How we run it
How the product is hosted
OrbySoft runs on Google Cloud. São Paulo is the home region. Agents use Gemini through Vertex AI.
Gemini and Gemma
What the agents call today, through Vertex AI. We can point at another model if the job needs it.
Cloud Run and GKE
Where functions and sandboxes run. They scale down when the channel is quiet.
BigQuery and Cloud Storage
Telemetry, evals, and versioned weights. One partner’s data does not train another’s model.
IAM and Secret Manager
Secrets stay out of the code. Each partner is a separate scope.
southamerica-east1
São Paulo by default. Another region only if you ask for it.
Subscription plus usage
A platform fee, then tokens and compute. The invoice can split by partner.
Notes
What we shipped recently
Agents scoped to one partner
Memory, tools, and policy stay inside that tenant. Even the logs.
Product
Sandbox snapshots
Freeze a short-lived machine and bring it back without a cold start speech.
Platform
São Paulo as the default
New environments land in southamerica-east1 unless you say otherwise.
DataStart with one partner
Tell us about the network. We’ll stand up an environment with a spend cap.