One endpoint in front of every model
An OpenAI-compatible endpoint on your domain, covering public providers and models on your own GPUs.
The auditable AI data boundary
One OpenAI-compatible endpoint in front of public providers and models running on your own GPUs, with a record of every governed crossing.
The endpoint is the delivery mechanism
The product is the boundary and the evidence that it held. The record answers who asked, through which key, for which model, where the request went, what came back and what it cost.
What ships now
Multi-provider routing
Prefix wildcards route new model identifiers without a configuration change.
Per-key identity
Every key carries consumer metadata, so calls do not arrive as anonymous shared traffic.
Full request and response audit trail
Body logging is selectable per consumer, and the record keeps identity, model, destination, token counts and cost.
Token-level cost accounting
Usage is priced against a maintained rate catalogue for major providers.
Kubernetes-native, offline delivery
A GitOps-friendly workload with an offline install bundle and a signed offline licence. There is no call-home check.
Subscription passthrough routing
The gateway routes supported subscriptions without holding the customer's provider credential.
What is planned
Turkish PII detection and redaction
Checksum-validated identifiers plus person and address detection in Turkish prompt text. Routing remains the compliance control.
Per-team limits, quotas and budgets
Gateway-enforced limits, spend ceilings and alerts per team.
Per-key model access control
Policy defining which keys may reach which models.
Local versus public cost comparison
The same workload priced on customer GPUs and on public providers.
Technical shape
- Interface
- OpenAI-compatible HTTP
- Runtime
- Kubernetes, GitOps-friendly
- Deployment
- Private, hybrid or public by workload
- Local inference
- Open-weight models through vLLM on customer GPUs
- Install
- Offline bundle with Helm chart and registry mirror
- Licence
- Signed offline file with an expiry, never a call-home check
- Owner and support
- DemirTech
Questions security teams ask
Does redaction make an overseas transfer compliant?
No. Masking and pseudonymisation do not remove data from KVKK scope when the link to a person can be restored. Routing is the control: a regulated workload stays inside the boundary. Redaction only reduces exposure for traffic that policy already permits to leave.
Can Infegate run without any internet connection?
Yes. The install bundle and licence are offline. Only open-weight models can run air-gapped, so the claim is well-operated open-weight inference, not access to closed frontier models.
What does GA mean here?
It applies to the seven shipped capabilities listed on this page, including the air-gapped delivery path. Planned compliance and control features remain explicitly outside that claim.