Provable boundaries for autonomous systems.
We build deterministic infrastructure for enterprise AI. DARM-Guard enforces machine-checked boundaries around autonomous agents. Vigil delivers real-time transaction surveillance for fraud prevention and anti-money laundering compliance. Aura orchestrates enterprise knowledge, connecting teams and AI agents to trusted, well-governed information.
Core Infrastructure Platforms
DARM & DARM-Guard
Machine-checked assurance calculus (DARM Core v1.0) & runtime broker (DARM Guard). Isolates agent proposals, checks compiled Lean 4 kernel, and records signed audit logs.
Vigil
Real-time fraud management & regulatory compliance solution. Evaluates transactions in sub-150ms with automated statutory STR/CTR returns.
Aura Platform
Enterprise knowledge base & multi-agent orchestration. Connects corporate documents to verified copilots with strict cryptographic audit logs.
Agency Hardening
Bespoke engineering engagements. We audit high-overhead enterprise workflows and deploy deterministic agent networks targeting 60% operational reduction.
Vigil: Pay-As-You-Go Surveillance
Engineered as a modular monolith on GCP. Scales to zero when idle with in-process ML inference and automated regulatory filings.
- Scoring: <150ms latency SLA
- ML Path: In-process serving, zero endpoint lock
- Compliance: Direct NFIU STR & CTR generation
Automation Agency Engagements
For enterprise organizations requiring rigorous process audits and customized boundary-hardened agent architectures.
Fixed scope audits with verifiable telemetry and risk scoring.
The Guard Boundary Pattern
Most modern AI architectures allow probabilistic language models to directly invoke APIs and trigger side effects. In enterprise environments — whether authorizing financial transactions or executing multi-step agentic workflows — this model is fundamentally flawed.
- GATE 01Model Ingest
- BOUNDARYDARM Broker & Kernel
- STATEPreserved & Proved
- Provenance Registry
- Lean 4 Decision Kernel
- Signed Audit Log
- Admitted Execution
- STAGE 01Raw IngestAPI request, transaction payload, or agent action proposal
- STAGE 03 · BOUNDARYDARM Guard ↗Lean kernel verdict & durable signed intent
- STAGE 04Governed EffectDeterministic commit, regulatory filing, or enterprise API call
Evidence, Bounds & Formal Constraints
- Unbounded Probabilistic AgentStandard ReAct or LangChain loopFAIL: INFERRED PRIVILEGESHallucinated tool arguments execute without static checks.
- WASM Perception SandboxLayer 2 memory-bounded sandboxPROVED: ZERO ESCAPEHard 512MB RAM ceiling; syscalls trap immediately.
- Vigil Sub-150ms Fraud PathIn-process ML & regulatory gateTESTED: 99.98% SLAEvaluates rules in <10ms, total pipeline <120ms median.
- DARM Guard Broker & Lean KernelBoundary-indexed admission & durable audit logLEAN 4 MACHINE-CHECKED31 claims, 53 effect sites, 12+4 closure checks.
# DARM Guard Broker v0.28.0 (commit 4bcab18)
# broker.py: Canonicalization -> Compiled Lean Kernel -> Durable Intent -> Effect
async def decide(self, request: InvocationRequest) -> AdmissionVerdict:
# 1. Canonicalize & freeze invocation before asking kernel
invocation = self.canonicalize(request)
# 2. KernelClient verifies verdict with compiled Lean kernel
verdict = await self.kernel_client.decide(invocation, nonce=request.nonce)
if not verdict.is_admitted:
<span class="diff">- return AdmissionVerdict.refuse(verdict.reason)</span>
# 3. Durable intent recorded to signed, hash-chained audit log
<span class="diff">+ intent = await self.audit_log.record_intent(invocation, verdict)</span>
<span class="diff">+ return await self.execute_admitted(invocation, intent)</span>
}Deploy deterministic intelligence to your edge.
Whether retrofitting 2,000 legacy RTSP video feeds across bank branches or locking down transaction scoring latency below 150ms, PerceptraAI provides audited, production-verified architectures.