Mandarum Praxon — Autonomy you can actually ship.
A runtime that executes agent graphs on NVIDIA GPUs with per-agent and per-action identity, policy-as-code and NeMo Guardrails checked before every model call or tool use, durable resumable execution, and an immutable audit log of every decision, prompt, and dollar.
Why most agent deployments are a liability.
Agents act with no oversight
An agent that can call tools and spend tokens is a liability without identity. Praxon issues a scoped identity per agent and per action.
Tools fire before policy checks
A prompt-injected agent shouldn't reach a payment API. Praxon evaluates policy and NeMo Guardrails before each step, not after.
'The agent did it' isn't an answer
Auditors need to see every decision. Praxon writes each prompt, tool call, guardrail verdict, and cost to a tamper-evident ledger.
What Praxon does.
Every agent and action is authenticated.
Scoped, short-lived identities per agent — and per action within a run — with capability and spend limits enforced in real time. Revoke an agent's access in one click.
- → Short-lived per-agent + per-action credentials
- → Capability + spend limits per agent
- → Instant fleet-wide revocation
policy agent-actions { when tool == "payments.refund" and amount > 50 require approval from role("ops") log reason }
Policy and guardrails before each step.
A declarative policy DSL plus NeMo Guardrails evaluate every model call and tool use at the perimeter — allowlists, escalation paths, and two-person rules on sensitive actions, in sub-millisecond time.
- → Sub-ms policy evaluation per step
- → NeMo Guardrails at the perimeter
- → Escalation + two-person rules
Durable execution with a full audit trail.
Agent runs are checkpointed and resumable — a failed step continues on another GPU exactly where it stopped. Every decision lands in a tamper-evident ledger exportable to your SIEM and mapped to SOC 2 and the EU AI Act.
- → Durable, resumable agent execution
- → Tamper-evident decision ledger
- → SIEM export · EU AI Act mapping
# 09:14:03 agent:support-triage step retrieve → guardrails_pass tool refund $42 → approved @ops run resumed → gpu-7 · logged
Praxon, in production.
One product. One platform.
Praxon governs the agents. Voltra serves the models they call. Cuvex grounds their recall in proprietary data. Edgeon carries the same identity and audit to Jetson edge nodes. Twinex models how policy changes affect fleet cost before rollout. Every product is stronger with the others.
Mandarum Voltra
Agentic scheduling and autoscaling that lifts GPU utilization and cuts $/token across H100/H200/Blackwell fleets.
Explore Voltra →
Mandarum Cuvex
GPU embedding, cuVS vector search, and reranking in one permission-aware recall path.
Explore Cuvex →
Mandarum Edgeon
Serve governed AI on NVIDIA Jetson fleets with cloud-parity policy, telemetry, and safe OTA.
Explore Edgeon →
Mandarum Twinex
Simulate GPU inference fleets on Omniverse to plan cost, capacity, and serving topology before production.
Explore Twinex →Ship agentic AI you can actually stand behind.
See how Praxon wraps your agents with identity, guardrails, and an immutable audit trail — without rewriting a line of agent code. Start free, or schedule a governance walkthrough.