Praxis
Ask for infrastructure. Keep control.
Provision and manage virtual infrastructure through plain-language requests — interpreted by AI, governed by your policies, approved by your people, and recorded end to end.
Six stages, and only one of them is the model.
Ask
A requester describes the outcome they need in ordinary language — no platform-specific form.
Interpret
The model reads the sentence and fills in a structured, reviewable spec. That is its only job.
Check policy
Deterministic rules allow, route for approval, or refuse — and say exactly which constraint applied.
Approve
Named, authorized people sign off when policy requires it. The AI never overrides them.
Provision
Approved work executes across platforms, with post-build profiles and verification.
Audit
A tamper-evident chain of every request, decision and action is retained.
Ask
A requester describes the outcome they need in ordinary language — no platform-specific form.
Interpret
The model reads the sentence and fills in a structured, reviewable spec. That is its only job.
Check policy
Deterministic rules allow, route for approval, or refuse — and say exactly which constraint applied.
Approve
Named, authorized people sign off when policy requires it. The AI never overrides them.
Provision
Approved work executes across platforms, with post-build profiles and verification.
Audit
A tamper-evident chain of every request, decision and action is retained.
One sentence in. A reviewable proposal out.
Describe the outcome, not the platform mechanics. Praxis infers what it can, asks for what it genuinely cannot, and shows you exactly what it understood before anything is created.
“I need a small Ubuntu server for the payments development team with 2 vCPU and 4 GB RAM.”
No hypervisor settings. No template names. No ticket translation.
- Environment
- Development
- Operating system
- Ubuntu 22.04 LTS
- Image
- ubuntu-2204-hardened-v7
- Size
- 2 vCPU · 4 GB RAM
- Network
- dev-app-approved (vlan 412)
- Storage
- 40 GB · standard tier
- Target cluster
- hv-nonprod-02
- Application
- Payments API · dev
- Business owner
- K. Rahman · Paymentsasked
Praxis asks only for what is genuinely missing — then nothing happens until a person has seen it.
Then the rules decide — not the model.
Each rule is a policy your team wrote, evaluated against the state of the estate at that moment. A refusal is never a dead end: Praxis names the constraint and offers alternatives already checked against the same rules.
- CPU-COMMIT-80Committed CPU on target ≤ 80%hv-nonprod-02 at 61%
- SIZE-DEV-MAXDevelopment size ≤ 4 vCPU / 8 GBrequested 2 vCPU / 4 GB
- NET-APPROVEDNetwork must be environment-approvedvlan 412 approved
- APPROVAL-PRODProduction requires named approvernot applicable — development
Decided by rules, not by the model.
- Committed CPU above policy limithv-prod-01 at 87% — limit 80%
- Requested size exceeds threshold16 vCPU — production maximum is 8 vCPU
- 8 vCPU on hv-prod-01within size and capacity limits
- 16 vCPU on hv-dr-02capacity available · DR cluster permitted
Praxis does not just say no. It explains what will work instead.
Thirteen minutes, nine seconds — and eleven of them were waiting for a human.
Praxis is fast where it should be and deliberately slow where a person has to sign. The gap between 09:41:04 and 09:52:18 is the approval, and it is the point.
- Intent received09:41:02Development Team
- Plan prepared09:41:04Praxis
- Confirmed09:52:18Infrastructure Governance
- Executed09:53:04Praxis Engine
- Verified09:54:11Praxis Engine
Every step SHA-256 hash-linked — edit one row and every hash after it breaks.
Self-service that never bypasses governance.
Intent-based requests
Create infrastructure without platform-specific forms or consoles.
Intelligent placement
Match workloads to the right environment, capacity and policy.
Policy gate
Allow, route for approval or refuse — deterministically.
Explainable decisions
Every choice, block and permitted alternative is spelled out.
Automated profiles
Reusable post-build steps: domain join, packages, CMDB.
Governed day-2 ops
Resize, add storage, clone and snapshot under the same rules.
Infrastructure visibility
Clusters, VMs, utilization, requests and outcomes in one view.
Audit & accountability
A tamper-evident chain of every request, decision and action.
Three things that are true on day one.
Runs inside your environment
Praxis is deployed in your data centre. Infrastructure data is not sent to a public AI service.
Locally hosted language model
The model that interprets requests runs on your infrastructure, under your control, offline.
Role-based access, separation of duties
Who may request, who must approve, and who may execute are distinct — and enforced, not assumed.
AI proposes. Policy decides.
Humans approve.
The model interprets the request and proposes a configuration. Deterministic rules make the decision. Authorized people approve when required — the AI never overrides policy or provisions on its own. A refusal is never a dead end: Praxis names the constraint that blocked the request and offers alternatives already checked against the same rules.
- The model only fills in a form
- It never touches infrastructure
- Deterministic, inspectable rules decide
- Named humans approve
- Credentials stay centrally managed
- Every decision is recorded
Five guardrails checked on every request, automatically.
- CPU-COMMIT-80 · SIZE-DEV-MAX
Capacity & size
Committed CPU on the target and the maximum size per environment — the rules that stop a request landing on a cluster that cannot carry it.
- ubuntu-2204-hardened-v7
Approved images
Only hardened, catalogued images are eligible. A request for something outside the catalogue is refused, with the catalogue offered instead.
- NET-APPROVED · vlan 412
Network & namespaces
Networks and namespaces must be approved for the environment being requested. Public exposure is not something a request can grant itself.
- Per team · per environment
Cost thresholds
Spend and sizing caps per team, environment and cost centre, evaluated before provisioning rather than reconciled afterwards.
- APPROVAL-PROD
Approval routing
Which requests need a named human, and which human. Production requires a named approver, and a requester can never approve their own request.
The model runs on your GPU. Offline.
Praxis is deployed inside your data centre and the language model runs locally — and if you switch the AI off entirely, a deterministic fallback takes over.
- No cloud API — the model runs on your own hardware
- No VM names, topology or credentials sent anywhere
- Works with the AI switched off; a deterministic fallback takes over
- Inventory access is read-only — Praxis cannot mutate what it reads
What Praxis will not do.
The limits are as deliberate as the features, and they are enforced in code rather than promised in a deployment guide.
- Delete, power off or detach anythingDecommission archives — it never deletes. Destructive operations are outside the product’s remit, not merely gated.
- Change networking, storage or cluster configInventory is read-only. Praxis reads the estate to reason about placement; it does not reconfigure it.
- Approve on your behalfA named human signs, and it is recorded. The AI never holds the final decision, and a requester cannot approve their own request.
- Run unapproved automationEditing a workflow sends it back to draft. Nothing executes from a profile that has not been approved in its current form.
Built for teams that need speed without losing control.
Platform & infrastructure teams
Self-service that never bypasses the guardrails they set.
Application & product teams
Operational requests answered in minutes, not tickets in a queue.
Security & compliance teams
Every request checked against policy, credentials kept out of end-user hands.
Organizations with data-residency limits
A locally hosted model, so infrastructure data never leaves the estate.
FinOps & cost owners
Spend and sizing caps enforced automatically, not audited after the fact.
Audit & risk
An answer to “who built this, who approved it, and what was done to it”.
Experience infrastructure on your terms.
Book a personalized Praxis demonstration — it runs inside your data centre, on your locally hosted model.