FIG. 11
Agent deployment
We deploy the stack, not just a chat window.
An agent, its memory, a model gateway, and the models you choose. Frontier models from
OpenAI, Anthropic, Google, or another provider are the default, called on your own
accounts. Open-weight models on your own hardware when data cannot leave the building.
We consult first, then build it, install it, and stay to run it.
01
Models: frontier providers by default
Agents call the strongest models available from OpenAI, Anthropic, Google, or
another provider, on your accounts, and each task is pointed at the model that
fits it. When data cannot leave the building, the same agent runs on open-weight
models on your own hardware instead.
02
One gateway in front of all of it
Every model call goes through an AI gateway we deploy and operate, so you get one
endpoint for every provider, a key per agent with its own spend cap, and failover
when a provider has a bad day. Swapping a model later is a configuration change,
not a rebuild.
03
Where the stack runs
Your hardware or private cloud, or a Gephyra Cloud instance: one dedicated machine
per client, hosted in the EU or the US, carrying the agent, its memory, and your
wiki. Nothing is shared with another customer, and provisioning starts the moment
payment clears.
FIG. 12
The stack we deploy
-
01
Agent. Hermes, the open-source agent runtime that does the
work: reading your documents, calling your systems, finishing multi-step tasks.
-
02
Memory. Hindsight, a long-term memory service, so the agent
keeps context between sessions instead of starting from zero every morning.
-
03
Gateway. LiteLLM, the AI gateway in front of every provider:
one endpoint, a key per agent, spend caps, and failover.
-
04
Coding agents. Claude Code and peers on development-heavy
work, so integrations get built rather than clicked together.
-
05
Wiki. Outline, a private knowledge base your team and the
agent both read and write.
-
06
Models. Your provider accounts (OpenAI, Anthropic, Google,
and others), or open-weight models running on your own hardware.
FIG. 13
What the gateway buys you
→
One endpoint
Agents are written against a single API, so adding a provider or changing models
later is a configuration change, not a rewrite.
→
Your accounts, your keys
Model calls run on your provider accounts. Inference is billed by your provider,
not marked up by us, and no key is ever resold.
→
Caps that hold
Every agent, team, or project gets its own key with a monthly budget. A runaway
loop hits the cap instead of your card.
→
Visibility and failover
Every call is logged with tokens and cost per key, and a provider outage fails
over to another model instead of stopping the workflow.
FIG. 14
What an engagement covers
-
01
Consult. A diagnostic that maps where AI fits, what it costs,
and what to build first. Written roadmap, fixed fee, yours to keep.
-
02
Build. The agent itself: your documents, your systems, your
workflow, tested against real cases before handover.
-
03
Deploy. Hardware, network, backups, monitoring, and the keys.
We install it, not only configure it.
-
04
Operate. Model updates, tuning, monitoring, and a named point
of contact, on a monthly subscription.
Not sure which models, or where it should run?
That is what the diagnostic is for. Most engagements start with one workflow, one
model choice, and a baseline we agree on up front, then grow from there.
If a frontier model on your account is the right answer we will say so. If local
hardware is the right answer, we will say that instead.
First step
A diagnostic, scoped and fixed fee.
Where it runs
Your hardware or a dedicated hosted box. Models on your accounts.