Services
Start from the infrastructure you have and the way it runs today. The platform is the foundation, and private AI is what we build on top. An engagement can be an architecture review, an on-premises pilot or a production cluster, with a clear quotation before work begins.
Each can be a standalone engagement or a phase of a full delivery. Platform work comes first because everything else sits on it. We agree the technical scope and send an itemised quotation before implementation begins.
What we deliver Four services on one platform, from the cluster to the decision a person still has to make.
Kubernetes and GPU platform. Clusters that schedule GPUs and survive a failed node. OpenShift, Nutanix NKP or upstream Kubernetes, on bare metal or air-gapped.
Private AI. Models served inside your network. Nothing leaves. Open-weight models on vLLM or Triton, sized to the GPUs you own.
Knowledge and authority. The right source, for the right people. Which documents are authoritative, and who may read them, is settled before anything is indexed.
AI governance. An agent can ask. A person decides. Anything outside the agreed scope stops until someone accountable approves it.
Scoped, built and handed over to your team.
We design and deliver the container platform your workloads need: Red Hat OpenShift, Nutanix NKP or upstream Kubernetes, with GPU scheduling and virtualisation. Built for bare metal, air-gapped and sovereign operation.
Open-weight and fine-tuned models served inside your data centre on high-throughput engines such as vLLM and Triton. No dependency on a cloud API, and no data leaving your network.
Make your organisation's information safe and useful for AI: which sources are authoritative, access boundaries that follow your existing roles, and a lifecycle that keeps documents current.
Decide what AI may do on its own and when a person must approve the next step. We design the controls around those decisions, including access, review, refusal and an audit trail.
Every engagement is broken into four concrete stages. Each stage produces a tangible deliverable you own, and each can be purchased standalone so you never commit blind budget.
We audit your existing hardware, data readiness, security constraints and workflow bottlenecks before you buy GPUs or licenses. Deliverable: written current-state assessment, target architecture, risk map and cost envelope.
We specify the exact cluster topology, local model serving stack, document authority rules and human-approval thresholds. Deliverable: reference configuration, hardware BOM sizing, acceptance criteria and rollback plan.
We deploy a bounded working system inside your environment running against your real Thai documents, LINE OA flow or GPU nodes. Deliverable: working on-premises pilot with measured accuracy and governance evidence.
We cut over to production, connect SAP, Express or FlowAccount ERP rails, train your operators and engineers, and hand over the Day-2 runbooks so your team owns the system.
Open the 8-slide Light executive sales pitch Read the Light AI governance & ROI presentation See full commercial catalog and how to buy Try the live Private AI workbench
This is what the AI governance service produces: a decision made before the action runs, not a rule written into a prompt and hoped for. The illustration below is the shape of that control, not a product you buy.
An illustration that runs in your browser and is connected to nothing. Start the request and watch where it stops. It will not go any further until you approve that exact revision yourself.
workflow-agent-17staging/*Operation: restart
Your approval covers only the target, operation and revision shown above.
DENIED · production/payment-api sits outside staging/*, and nobody has approved it.
Illustrative evidence: identity workflow-agent-17, the target and revision, the policy decision, and what the approval actually covered. After a real action you still compare what happened against what was approved, and reconcile any difference.
A restart of production/payment-api by workflow-agent-17, which is authorized only for staging/*, is refused. A human approval naming that target, the restart operation and revision 182 permits that one request and nothing else. Identity and permissions come from the execution layer, never from what the model claims about itself.
AI handles the volume; people handle the judgment. What gets escalated is written down as a rule during design, so nobody is asked to trust a confidence score without seeing the evidence behind it.
Agents can run on local, enterprise-hosted or frontier models. Identity, authorization, scope, approval, tool access, state, audit and reconciliation are checked before any of them reach your systems. Choosing a different model does not change who is allowed to do what.
Your environment. Clear boundaries.
Reasoning can be flexible. Actions need permission.
Review the corrections people made, the outputs they rejected, the escalations and exceptions, how long things take and what changed in the data and the business rules. AI can propose improvements, but every change is tested, reviewed and released as a controlled revision. Production policy never changes quietly. Ongoing support is agreed separately, per project.
Tell us what you want to achieve and what you run today. We scope the work and send an itemised quotation within 48 hours.