System
A kernel-first system for physical agents.
Keplen’s system design starts below the model and above the actuator.
The goal is a governed physical AI stack where agents can perceive, plan, delegate, and act only through explicit identity, scoped authority, bounded actions, state commits, and traceable evidence.
Stack
Five layers, one governed center.
Each layer has a clear owner and a clear boundary. Enforcement stays in the runtime; everything above it proposes, and everything below it executes only what the runtime allows.
- 01
Signal layer
Text, code, audio, visual/video, sensor streams, state references, and future action traces become structured signal inputs.
- 02
World-model layer
A signal-latent world model learns shared state, prediction, uncertainty, and proposal behavior across modalities.
- 03
Runtime layer
Splendor Kernel provides the governed loop: identity, work orders, action gateway, verifiers, quotas, state commits, traces, replay, and governance hooks.
- 04
Coordination layer
Vaisys Harmony provides private AI foundation and digital-agent tooling. Harmony-style control planes can issue scoped work orders, approvals, and artifact references while runtime enforcement remains with Splendor.
- 05
Alignment infrastructure
ValueAI supports the open alignment layer: datasets, model artifacts, training recipes, algorithms, evaluators, provenance, and release discipline.
Fleet coordination
Fleet coordination should not begin with dashboards. It should begin with authority.
Keplen’s fleet path is based on explicit node identity, capabilities, heartbeats, scoped work placement, local policy caches, trace synchronization, state handoff, and telemetry that informs but does not silently authorize.
- Node identity
- Capability declaration
- Heartbeat
- Scoped work order
- Placement policy
- Local validation
- Trace synchronization
- State handoff
- Telemetry
- Human intervention path
Physical boundaries
Physical actions must stay high-level and bounded.
A Keplen-compatible physical agent should request actions such as inspect zone, move to waypoint, return to base, dock, or read battery. It should not receive raw actuator authority from a cloud planner, chat interface, or unverified digital agent.
Allowed direction
Model proposes → Runtime verifies → Adapter executes bounded action → State commits → Trace records outcome
Rejected direction
Remote model → Direct actuator command
Self-management
Self-management does not mean unchecked autonomy.
For Keplen, self-management means an agent can maintain state, request work, report uncertainty, pause, escalate, recover from partial failure, and preserve evidence for review.
- Persistent identity
- State awareness
- Policy validity checks
- Quota awareness
- Uncertainty reporting
- No-op behavior
- Escalation path
- Replayable traces
- Recovery mode
- Human review hooks
Integration model
Who owns what.
Keplen combines independent systems rather than collapsing them. The boundaries between them are the point.
Splendor Kernel
Runtime enforcement, identity, action gateway, verifiers, traces, replay, state commits, governance hooks.
Vaisys Harmony
Private AI foundation, digital-agent tooling, control-plane workflows, enterprise integration path.
ValueAI
Alignment datasets, model artifacts, training recipes, algorithms, evaluations, provenance, release discipline.
Keplen
Physical AI research company combining world models, runtime boundaries, self-management, fleet coordination, and alignment process.
Build with Keplen.
We are looking for teams working on edge autonomy, robotics middleware, runtime verification, digital-agent tooling, fleet operations, simulation, evaluation, and alignment infrastructure.