RiskPilot · Capability brief
An AI-augmented cybersecurity risk-management platform — engineered three ways, with full transparency.
A complete NCA-aligned risk-management system with an AI Risk Scenario Generator at its core. Three architecture paths — frontier-AI capability, private single-tenant deployment, full air-gap isolation — pick by what matters most: speed, control, or regulatory tier.
منصة تفاعلية لإدارة مخاطر الأمن السيبراني، متوافقة مع الهيئة الوطنية للأمن السيبراني، بثلاث معماريات معتمدة وتسعير شفّاف وعرض حيّ للذكاء الاصطناعي.
What RiskPilot delivers
Built around a standard NCA risk-management workflow — 15 steps, 5 frameworks, AI augmentation end-to-end.
The scope, in one paragraph
An automated cybersecurity risk-management system covering the full NCA control suite — ECC, CCC, CSCC, DCC, OSMACC — across 15 workflow steps from trigger through continuous reassessment. Risk-matrix scoring, SLA-driven treatment plans, dashboards, Power BI reporting, audit trail, and an AI Risk Scenario Generator at the centre.
The privacy property
- Security + privacy — no data leakage, no model training on client data.
- Air-gap is one option, not the only path — privacy is a property, not a topology.
- NCA-aligned compliance content, mapped from official control catalogues — not generic templates.
- AI does the heavy lifting on scenario generation and compliance mapping; humans review every commit.
The reframe: privacy is a property achievable three different ways, with different cost-and-capability tradeoffs. Air-gap is one of them — not the only one. The three paths are laid out below, side-by-side, with a live demo so the system can be felt rather than imagined.
Architecture · the three paths
Each path gets you the same system. The difference is where the AI lives and how privacy is guaranteed.
Hybrid with Frontier AI
هجين مع نماذج الذكاء الاصطناعي المتقدمة
The system runs on a single-tenant server the client controls. AI calls (Risk Scenario Generator + Compliance Mapping Engine) route to Anthropic Claude or OpenAI GPT under an enterprise contract with zero data retention and no training on inputs. All other data — risks, controls, audit log, evidence — stays on the client's server.
Privacy posture
Contractual + technical isolation. Enterprise no-train DPA, encrypted in transit, zero-retention policy.
Capability ceiling
Frontier (Claude Opus 4.7 / GPT-5 class). Substantially better risk-scenario quality than any open-source model in 2026.
Best for
Teams that want frontier-grade AI and the fastest path to a working system, with a contractually-isolated cloud AI provider.
The tradeoff
AI inference data leaves the client network for milliseconds. Enterprise contracts forbid retention and training, with audit rights. Acceptable for most regulated workloads.
Components
- Application serverNext.js + Node 22, single-tenant on a Saudi/UAE cloud region or a private VPS
- DatabasePostgreSQL 16 + pgvector, encrypted at rest
- AI providerAnthropic Claude Sonnet 4.6 (primary) · OpenAI GPT-5 (fallback) · both under enterprise zero-retention DPA
- AuthenticationLocal users + Entra ID / SAML option
- DashboardsPower BI Service OR self-hosted dashboards (Power BI Service is allowed on this path)
- Email alertsInternal SMTP or SendGrid/SES
- Audit logAppend-only Postgres table, 7-year retention default
One-time
$7,000–14,500
Time
2–5 wk
Monthly
$670–830
No GPU required. AI runs at the vendor under enterprise contract.
Side-by-side
The full comparison — no salesy obfuscation.
Every dimension that actually drives the decision, in one table. The cell with the stronger position is highlighted.
| Dimension | Path C · Hybrid | Path B · Private | Path A · Air-Gapped |
|---|---|---|---|
| One-time build (all phases) | $7K–14.5K | $10.5K–19.5K | $13.5K–25K |
| Time to ship v1 | 2–3 weeks | 3–4 weeks | 4–5 weeks |
| GPU hardware capex (client side) | $0 | $7K–35K | $28K–35K |
| Monthly run-rate (after launch) | ~$670–830/mo | ~$1,200–3,200/mo | ~$2,000/mo |
| AI capability ceiling | Frontier (Claude/GPT) | Open-source (Qwen 72B) | Open-source (Qwen 72B) |
| Risk-scenario latency (single user) | 1–3 sec | 2–5 sec | 2–5 sec |
| Privacy posture | Contractual no-train + zero-retention DPA | Cryptographic + architectural | Physical air gap (verifiable) |
| NCA tier fit | Essential / Advanced (with cloud DPA review) | Essential / Advanced / Minimal | All tiers incl. Critical National Infrastructure |
Highlighted cell = the path that wins on that dimension. No single path wins on everything — that's why we're showing all three. Pick by what matters most to you.
Live demo · the AI Risk Scenario Generator
Describe any asset. Four streams launch in parallel — asset profile, three distinct scenarios (insider / external / drift), then a synthesis call.
Real Claude calls, streamed token-by-token. The skeletons fill in as the model thinks. Each run is a fresh analysis — no canned data.
The 15-step workflow
Every step, with the artifact it produces.
Click a step to see what it does. The system enforces ordering, but a human reviews every AI-produced artifact before commit.
Step 1 · what it does
Trigger / Risk Assessment Initiation
إطلاق التقييم
Triggered by new projects, changes, cloud services, third parties, incidents, or vulnerabilities. Creates a Risk Assessment Ticket linked to the asset or project.
Artifact produced
Risk Assessment Ticket
Risk-level matrix
Likelihood × Impact = Risk Score
Both axes on a 1–5 scale. Cell values are the score; scale below.
| Impact 1 | Impact 2 | Impact 3 | Impact 4 | Impact 5 | |
|---|---|---|---|---|---|
| L=5 | 5 | 10 | 15 | 20 | 25 |
| L=4 | 4 | 8 | 12 | 16 | 20 |
| L=3 | 3 | 6 | 9 | 12 | 15 |
| L=2 | 2 | 4 | 6 | 8 | 10 |
| L=1 | 1 | 2 | 3 | 4 | 5 |
Treatment SLA grid
Risk level → required response time
Auto-applied on every treatment-plan record.
- Very LowLight monitoring—
- LowPeriodic review—
- MediumPlan within 20 business daysExecute within 6 months
- HighPlan within 5 business daysExecute within 3 months
- CatastrophicPlan within 5 business daysExecute within 1 month
Security & privacy
Three different mechanisms. Same property: client data stays inside the client perimeter, and is never used to train any model.
When the real concern is privacy + no leakage + no training on inputs — not the air gap itself — there are three legitimate ways to deliver that property. Each path explained below.
Path C
Contractual + technical isolation. Enterprise no-train DPA, encrypted in transit, zero-retention policy.
- Anthropic Enterprise: zero data retention by default, no model training on customer inputs (DPA Article 3), SOC 2 Type II + ISO 27001 certified
- TLS 1.3 for every API call — no plaintext on the wire
- Client data: stays in the single-tenant server. The AI provider sees only the specific question being asked, never the full database
- Audit rights: quarterly DPA compliance reports available from Anthropic on request
- Kill switch: fall back to Path B (self-hosted Qwen) at any time without re-architecting — point the AI client at a local endpoint
Path B
Cryptographic + architectural isolation. Self-hosted model, client-controlled server, client-defined network rules.
- Model + data on the same machine inside the client network: AI inference never crosses the perimeter
- Inputs and outputs processed in memory; nothing persists outside Postgres
- Network: VPN-only access · IP whitelist · TLS 1.3 · per-user audit logging on every read/write
- Encryption at rest: full-disk LUKS encryption + Postgres TDE on sensitive columns
- Hardware: a single rack box in the client DC or a dedicated cloud instance (AWS me-central-1 Saudi region, or Saudi Cloud)
Path A
Physical isolation. Verifiable via packet capture — zero egress traffic possible.
- Physical air gap: no Ethernet, no Wi-Fi, no Bluetooth, no cellular. The box cannot phone home because no path exists.
- Verifiable by packet capture on the upstream switch — zero egress traffic possible
- Model weights, vectors, audit logs, and all data persist only on this single box (with encrypted backup to a second internal box)
- All external integrations replaced by on-prem equivalents (e.g., internal SIEM only, no hosted Threat-Intel feeds)
- NCA Compliance Audit Pack recommended for this path — control-by-control documentation + external auditor coordination
On the no-training contract (Path C)
What "enterprise zero-retention" actually means.
- Anthropic enterprise terms: prompt + completion data deleted after request processing. No retention unless explicitly opted in for abuse-monitoring.
- No training: enterprise customer data is explicitly excluded from model training data, contractually guaranteed in the DPA.
- Compliance certifications: SOC 2 Type II, ISO 27001, HIPAA-eligible, GDPR-compliant.
- Audit rights: compliance reports and the DPA itself available before signing.
- Data residency: Anthropic offers EU-region routing; OpenAI offers region pinning. The closest jurisdiction to the client's residency needs is selected during scoping.
- Kill switch: if regulators shift posture, the system falls back to Path B (self-hosted Qwen) by changing one environment variable. Zero re-architecture cost.
Pricing · phase by phase
Transparent. Phased. Competitive against Big-4 consulting alternatives for comparable scope.
Phased delivery — clients stop at the end of any phase, no scope lock-in. Phase 0 (paid scoping) credits 100% against Phase 1 on proceed.
| Phase | Time | Path C | Path B | Path A |
|---|---|---|---|---|
Paid Scoping (optional) التحديد المدفوع Written technical spec, architecture diagram, firm fixed-fee quote. Credited 100% against Phase 1 on proceed. | 1 wk | $500 | $500 | $500 |
Phase 1 — Core v1 المرحلة الأولى — النواة All 15 workflow steps · AI Risk Scenario Generator · ECC + CCC compliance mapping (RAG) · Risk Register · Dashboards · Audit log · Email alerts · Mocked external connectors for SIEM/SOC/Scanner/Threat-Intel · Single org · Up to 10 users. | 2–4 wk | $7,000 | $10,500 | $13,500 |
Phase 2 — Frameworks Expansion المرحلة الثانية — توسعة الأطر CSCC + DCC + OSMACC mapping · Multi-organization support · Evidence Repository (file upload + indexing) · Enhanced reporting + custom report builder. | 1–2 wk | $4,000 | $5,000 | $6,000 |
Phase 3 — Live Integrations المرحلة الثالثة — التكاملات الحية Replace mocked connectors with real integrations to SIEM, SOC, Vulnerability Scanner, and Threat-Intelligence feeds. Continuous Monitoring pollers. Real-time alert routing. | 1–3 wk | $3,500 | $4,000 | $5,500 |
NCA Compliance Audit Pack (optional) حزمة التدقيق وفق NCA Control-by-control mapping document · audit-trail templates · external auditor coordination · evidence-collection runbooks. Recommended for Path A; optional for B and C. | 1 wk | $2,500 | $3,000 | $3,500 |
| Phases 0 + 1 + 2 + 3 total (excluding optional audit pack) | 5–10 wk | $15,000 | $20,000 | $25,500 |
Monthly run-rate
After v1 ships
| Line item | C | B | A |
|---|---|---|---|
| AI inference (Claude API · enterprise zero-retention) | $20–80/mo (200–1,000 assessments) | $0 | $0 |
| Hosting infra (single-tenant, Saudi/UAE region) | $50–150/mo | $0–2,000/mo (self vs cloud GPU) | $0 (client facility) |
| Maintenance + support (99% uptime · 4h P1 · 1-bday P2) | $600/mo | $1,200/mo | $2,000/mo |
Maintenance billed quarterly in advance, cancellable with 30 days' notice.
GPU hardware (Paths B + A)
Sourced via Saudi distributor, no margin
Hardware procured by the client directly. Spec and sourcing support included — no markup, no inventory risk on either side.
- 1× NVIDIA H100 80GB PCIe$25–28KRecommended for Path B (good balance) · paths B
- 1× NVIDIA H100 80GB SXM$28–32KRequired for Path A (enterprise air-gap grade) · paths A
- 2× NVIDIA A100 80GB$15–20K used / $30–35K newMost common, well-tested with vLLM · paths B / A
- 1× NVIDIA L40S 48GB$7–8KEntry tier; supports Qwen 32B only (not 72B) · paths B
Notes & assumptions
- All prices in USD. SAR conversion at 3.75:1 (pegged).
- Billing schedule: 40% kickoff · 30% Phase 1 acceptance · 30% Phase 2 acceptance. Phase 3 billed separately upon scope sign-off.
- Maintenance + support billed quarterly in advance, cancellable with 30 days' notice.
- Not included: hardware capex (Paths A & B), AI API consumption (Path C, billed direct to the client by Anthropic/OpenAI), existing software licenses (Power BI, OS, etc.), Saudi WHT (deducted by the client at payment per ZATCA).
- Engagement led by a senior AI engineer with hands-on production experience across compliance-adjacent platforms.
Delivery timeline
Side-by-side, week by week.
Each row is a path. Each block is a phase. Engagements begin with Phase 0 (paid scoping) upon path confirmation.
About
Engineered hands-on by a senior AI engineer with shipped production AI systems.
RiskPilot engagements are led by Omar G. Nagy, an AI engineer with two years of shipping production AI systems for compliance-adjacent businesses. Hands-on engineering throughout the engagement — no handoff to a junior team — with fast iteration cycles and honest tradeoff communication, including when the honest answer cuts against the higher-priced option.
This interactive proposal site — including the live AI demo, the architecture diagrams, and the pricing transparency — was built and deployed in under 24 hours. The same engineering velocity carries into every engagement.
Selected production work
- RetailOSAI-first multi-tenant retail SaaS. Production. Live customers.
- MedPrüfAustrian medical-exam prep platform — 10K+ AI-augmented questions, GA4 + Clarity instrumented.
- NeuraScaleFirst-party analytics + booking + lead-qualification platform; own-stack.
- Bridge SourcingB2B sourcing platform leveraging Egypt's 0% tariff lanes.
More at omargnagy.com
Next step
Choose a path — or start with paid scoping.
Upon path confirmation, a fixed-fee SOW is delivered within 24 hours. When a client is between two paths, a one-week paid scoping engagement ($500, fully credited against Phase 1) produces a written technical spec and a firm quote with zero ambiguity.
بعد تأكيد المسار، يُسلَّم عقد SOW بسعر ثابت خلال 24 ساعة. أسبوع التحديد المدفوع ($500، يُخصم بالكامل من المرحلة الأولى) يُنتج مواصفات تقنية مكتوبة وعرض سعر نهائي بدون أي غموض.