Service 01 — AI Gateway & Security
Your AI gateway. Your infrastructure. Your control.
We build custom AI gateways that sit between your applications and your AI models — a centralised layer for routing, security, governance, monitoring and observability, with in-house PII and PHI detection running before sensitive data ever reaches a model.
Deployment
Your VPC or on-premises
Overview
An AI gateway is the difference between a governance policy and an enforced control. Once every model call passes through one gateway, you can enforce who reaches which model, detect and redact sensitive data before it crosses the trust boundary, attribute cost per team, and produce an audit record for every decision — all centrally, without touching application code.
We build these gateways for enterprises whose data cannot leave their perimeter. The detection models are ours and ours alone — trained on the data classes you actually process, rather than a generic regex pass that catches formatted identifiers and misses everything embedded in free text.
Capabilities
How we deliver
Detection that runs before dispatch, not after
Once sensitive data is inside a prompt, it has already crossed the trust boundary — and most providers will have logged it. Detection therefore has to happen in the request path, before the model is called, with redaction applied inline. Our detection models are trained on the data classes you actually handle, because the risk in regulated data is usually in unstructured text rather than in well-formatted identifiers.
Policy as a routed decision
Every request is evaluated against data class, model and environment, producing an allow, deny or redact decision that is logged with the request. That is what makes routing auditable: the answer to 'why did this request go there' is recorded at the moment it happened, rather than reconstructed from whichever client library a developer happened to import.
Routing, resilience and cost control
One endpoint dispatching across every model and provider you use, with failover when a provider rate-limits or degrades. Per-team budgets and rate limits live in the same layer, which turns the gateway from purely a security control into a financial one — the difference between knowing your AI bill and being able to reduce it.
Auditable governance by construction
Every request is logged with its policy decision, the redactions applied, the model selected and the outcome, retained to your compliance window. Because the audit trail is produced by the enforcement path itself, it is evidence you have rather than evidence you have to go and produce.
What you get
- Sensitive data is detected and handled before prompts leave your perimeter.
- All model traffic flows through one enforced path with consistent policy decisions.
- Every request carries an audit record sufficient to reconstruct the decision made.
- Cost is attributed per team and per feature, with enforceable budgets.
- Provider outages and rate limits no longer propagate to every dependent feature.
Frequently asked questions
What does a custom AI gateway provide?
A custom AI gateway acts as a secure, centralised proxy between your applications and AI models. It handles model routing, rate limiting and observability while running in-house PII and PHI detection and protection models to filter sensitive information before it reaches the models.
Why not just use a managed gateway?
Managed gateways are excellent when your data is non-sensitive and you use one provider. They become limiting when detection must match a specific regulated domain, when you run multiple providers or your own models, or when data must never leave your infrastructure — in those cases the policy model is fixed by the vendor rather than by your obligations.
Can it run entirely inside our infrastructure?
Yes. The gateway is deployed into your own cloud account or on-premises, with no outbound traffic to us. Detection models, policy and logs stay inside your perimeter.
How much latency does the gateway add?
Routing and metering add low single-digit milliseconds. Detection adds more depending on model size, and the engineering goal is keeping that below the noise floor of your end-to-end inference latency. We measure both rather than assume.
Does it work with our own fine-tuned models?
Yes. Self-hosted and fine-tuned endpoints are governed by the same policy model as commercial APIs, which is what keeps a single answer to the question of what is allowed.
Talk to an infrastructure architect
Bring your traffic profile, data constraints and cost targets. We will tell you what is feasible and what it would take.
Book a consultation