Enterprise AI gateway
KLIA applies identity, permissions and quotas to every request before routing it, and records who used which model and at what cost. Open models on your own infrastructure, European providers or external APIs: you pick the provider, you define the policy once.
KLIA
gatewayExposed API
Processing
Destination
Identity, permissions and quota before routing. What is denied never leaves.
Products
KLIA started out as a router. Today it is the base our vertical products run on — consuming the very same platform your teams do.
gateway Routing and control
The core. An OpenAI-compatible API for all your models, and a single place to decide who uses each one.
documents Documents into data
Intelligent document processing: classifies, extracts and reconciles the fields of case files, invoices and forms.
Intelligent Municipal Assistant
Conversational citizen support for town halls. Answers on procedures, services and tourism with the municipal portal sources in plain sight.
Platform
Order matters: nothing gets routed before we know who is asking, what they are allowed and how much they have left.
01
Personal token or scoped JWT. The person, their team and their organisation are resolved.
02
Models are enabled in a cascade: the organisation opens, the team opts in, the token narrows it down.
03
Consumption metered in real time. Limits per period, per person or per service.
04
To the chosen model: local on your GPU, a European provider or an external API. The same interface in all three cases.
05
Who, which model, how much and at what cost. Traces and metrics in OpenTelemetry, plus per-team reports ready to charge back.
OpenAI-API compatible. Opencode, Kilo, Cherry Studio, LibreChat, Jan or Lobechat work by changing only the base_url and the token. And OpenTelemetry-compatible: observability lands in your stack, not in yet another dashboard.
MCP hub · artifacts
KLIA does more than route inference. It is a hub for remote MCP servers — your own and third-party — governed per organisation and team, with a temporary artifact store in the middle.
MCP servers are registered once and exposed to teams with the same identity, the same permissions and the same quota as the models. Nobody is handing third-party server credentials around.
Chaining tools normally forces every intermediate result to travel back to the client and out to the server again. KLIA parks it in a temporary store: one tool drops the artifact, another picks it up, inside the perimeter.
This is what makes KLIA a platform rather than a proxy with metrics.
A modern OCR engine reads every character precisely, but it does not understand what it reads. A vision model understands the document, but it can invent a digit and state it with complete composure. KLIA cai runs both in parallel, cross-checks field by field what each one says, and layers deterministic validators and rules tailored to your business on top. Three layers correcting one another, inside your network and with no token meter in the middle.
Input
PDF · JPG · PNG · TIFF · scanned · multi-doc bundle
Output · structured data
No per-document billing. Cost does not grow with volume, because it runs on your own hardware.
100 % of the data lifecycle happens within your infrastructure. Neither the document nor what is extracted from it goes out to external services.
A new document type is live in hours with few-shot learning, from a handful of examples and without retraining models.
Throughput depends on the document type, its complexity and the number of fields to extract. Scale by adding GPU or by distributing the services across nodes, and export the trace of every run in OpenTelemetry.
AMI · Intelligent Municipal Assistant
A conversational assistant for town halls. It resolves questions on procedures, services, tourism and citizen support using only what the town hall itself has published — and showing where each answer came from.
It answers from the real content of the municipal portal and its related sites, such as the tourism portal or the e-government portal. If something is not published, it says so and points to the support channel instead of inventing a deadline or a fee.
Every answer carries links to the pages consulted. Citizens can check it and the town hall can audit it.
The language set is configurable: Castilian, English, Spain’s co-official languages —Catalan, Galician, Basque— or any other. Always over the same content, without maintaining one knowledge base per language.
Token consumption is included in the price, with no volume caps. The only limit is fair use, there to stop abuse and automated attacks.
Public demo Villa Laude Town Hall We built an entire made-up town hall — complete with its Dormant Volcano Spa, its Lake of the Flying Fish and its Museum of Impossible Objects — so you can try AMI on real municipal content without asking any town hall for its data.
Deployment
You choose where KLIA runs, and which providers it may reach.
KLIA deploys into the same cloud you already operate in, and keeps calling the providers you already have under contract. It does not replace them: it decides what each team may ask of them.
helm install klia klia/klia
A machine outside your cloud, with your usual providers and, if it suits you, local models on their own GPU for whatever you would rather not send out.
docker compose up -d
No route out, local inference only. For your most sensitive work — the kind that admits no external provider under any circumstances.
providers: local-only
KLIA adds no compliance surface. It runs inside the perimeter you have already accredited, with your identity and your policies. KLIA is not one more third party to audit.
Who is behind it
LAUDE is a software consultancy specialising in applied artificial intelligence: structured, supervised, measured AI woven into how the organisation actually operates. KLIA, KLIA cai and AMI are our own products, and they can be tailored through LAUDE’s professional services. Not everything is configurable, but it is worth asking.
LAUDE is accredited at ENS medium level. It applies to the services LAUDE operates — such as KLIA cai in its managed form.
The product is deployed and configured. The effort goes into integrating with your systems and defining your document types.
Let’s talk
And if what you need is something else, we’ll tell you that too.
Loading the form…