Smart Protection.
Full-spectrum AI governance.

SecuAgent is an enterprise AI security gateway built for organizations that need to safely handle personal data and trade secrets between cloud LLMs and on-premises systems. Its core capabilities are edge-based data redaction, intelligent request routing, and hybrid cloud/on-premises compute scheduling — letting enterprises adopt AI safely on a predictable budget.

The redaction engine combines regex pattern matching with lightweight NLP to detect and mask personal-data signatures in real time. The guardrail layer adds cross-validation between SecuGuard (F1=0.9653, 0.19% false-positive rate) and IBM Granite Guardian 3.0, reducing the risk of any single model's blind spots. We recommend using the 14-day free assessment to tune and validate against your actual conversation patterns.

No. The knowledge base uses multi-RAG partitioning, split into independent partitions by department and sensitivity level, so access rights directly define the retrieval boundary rather than filtering results after the fact. This is paired with four-dimensional RBAC/ABAC authorization (department, role, clearance level, need-to-know) — if any authorization attribute is missing, the system defaults to "access denied" (fail-closed), so a misconfiguration never leaves you exposed.

We offer a 14-day free assessment, with our engineering team guiding you from PoC to production. Endpoint protection intercepts transparently at the system, browser, and CLI layers, so employees don't need to change their existing habits. A third-party penetration test is completed before go-live.

We use a hard quota circuit-breaker plus a flat monthly rate, avoiding the "bill shock" of usage-based cloud API billing. Actual pricing depends on your usage volume and deployment model (on-premises or hybrid cloud) — book an assessment for a concrete quote.

Not yet. ISO/IEC 27001, ISO/IEC 42001, and ISO/IEC 27701 are all part of our planned roadmap, while SSDLC/SBOM and OWASP ASVS/penetration testing/vulnerability scanning are already integrated into our engineering process — we clearly label planned items as "planned," not obtained, and will update this page as each is achieved.

We resolve this with crypto-shredding: every audit event is chained with hashes, so any tampering is immediately detectable, satisfying long-term retention and non-repudiation requirements. A deletion request removes only that data subject's unique decryption key — the historical trail stays fully intact, but the personal data itself can never be recovered. If litigation or an investigation is underway, deletion requests are also automatically frozen by our legal hold mechanism.

The three-path interception architecture (system-level TUN/WFP, browser extension, CLI watcher) is designed for company-issued, managed devices, browsers, and dev environments. Fully unmanaged personal devices fall outside standard coverage — if BYOD is part of your environment, we recommend raising it during the assessment.

No vendor lock-in. Smart routing dispatches simple queries to lightweight open-source models and complex reasoning to top-tier commercial models, with millisecond-level automatic failover, and the model deployment layer exposes a unified API wrapper.

Admission Control immediately returns 429 + Retry-After for over-quota requests instead of queuing indefinitely. Human review is only triggered by high-risk actions — external sends, data changes, or highly sensitive content — so ordinary queries are never held up; overdue reviews escalate automatically at T+0/4h/8h/24h. Our engineering team helps investigate false positives and tune rules during onboarding, and for formal SLAs and support response times, we recommend agreeing on specifics with our team during the assessment.

SecuAgent is built specifically for the "low-cost, high-privacy, leak-proof, governable" needs of Taiwan's 1.6 million SMEs, not just large enterprises — the architecture scales resource allocation to tenant size. We recommend a free assessment to confirm the actual complexity and cost fit your situation.

The GPU utilization gain comes from vLLM PagedAttention memory management (from roughly 40% before deployment to 80%+ after). The egress cost reduction comes from JuiceFS local data caching (≥10GB/s NVMe reads, <0.1ms latency), which cuts cross-cloud data transfer. These are internal measurements — actual results will vary with your existing infrastructure and usage patterns.

The guardrail layer intercepts 14 common categories of security threats in real time, capturing and backing up security event logs for subsequent forensic analysis and compliance audits; every deployment passes third-party penetration testing before go-live. Because the system is deployed on-premises, the actual incident-reporting process and liability boundary (e.g., infrastructure security vs. software service) will be clearly defined in the contract and during onboarding.

Yes. SecuAgent automatically optimizes prompts at the edge before they reach the cloud AI — adding clear structure, explicit instructions, and response format constraints — improving output quality across the workforce while cutting wasted compute.

Yes. SecuAgent anonymously analyzes the frequency, complexity, and quality of employee-AI interactions and produces an "AI Competency Matrix" report, giving HR and IT leaders a clear view of company-wide AI proficiency to guide training and measure transformation progress; any small sample group that could be re-identified is automatically merged up to the department level (k-anonymity) rather than simply hidden, and the analysis never reconstructs what any individual actually asked — so management gets real insight without compromising employee privacy.