Project Keystone · Architecture & Security Reference
Nothing hidden. Engineered to be inspected.
Whether your CTO, Chief Architect, or CISO is reviewing Project Keystone, the entire design is explicit: seven clear layers, mechanical security boundaries, strict run accounting, universal MCP connectivity, and AI that builds dashboards, reports, and automations safely.
The 7-Layer Stack
How data rises from raw sources to governed AI, dashboards & automations
Select any layer below to inspect its role and security boundary:
Experience & AI Builder Layer
Natural speech Q&A, AI-built dashboards, custom reports, automations & operator console
Talk to your data in natural speech using the built-in LLM, instruct AI to build live dashboards, custom reports, and automations, or inspect the interactive Data Map and Operator Console (job health, freshness, exception ledger, access admin).
What happens here during a query or automation:
01 · You ask a question in natural speech, trigger an AI automation, or connect your LLM via MCP.
End-to-End Request Trail
- →01 · You ask a question in natural speech, trigger an AI automation, or connect your LLM via MCP.
- →02 · Keystone verifies who is asking (user or MCP client) and what their RBAC role and PII grants allow.
- →03 · One agreed business definition is applied—the exact same calculation every dashboard, report, and automation uses.
- →04 · The answer, dashboard, or automation trigger is assembled from clean, identity-resolved business models.
- →05 · Every model row links back to immutable raw storage kept under your own encryption keys.
- →06 · Every row was verified and accounted for on arrival, with full audit and freshness trails.
- →07 · The verified result, dashboard, or automation executes with a complete lineage trail back to the source.
Enforced by Code & Database Constraints
Six architectural rules that never rely on human memory
Every row is accounted for
rows_read = inserted + skipped + excepted
Checked both in code and by a database constraint on every job run. Bronze is strictly append-only—nothing is silently dropped or overwritten in place.
One definition, everywhere
Semantic compiler refuses duplicate or ambiguous metrics
Metrics, units, currencies, and additivity rules are declared once. Every dashboard, natural-speech AI query, MCP client, and automation uses the exact same definition.
Any LLM via MCP, zero hallucinated numbers
Built-in LLM + OpenAI, Claude, Gemini & In-House LLMs
Connect any chosen LLM via MCP or use the built-in natural-speech AI to ask questions and build dashboards, reports, and automations—with every figure verified against the semantic layer.
Arrival time never decides current state
Explicit source ordering & epoch reset detection
Selecting the latest row by when it landed breaks on every replay or backfill. Every entity declares its own ordering rule, tie-breaker, and upstream reset detection.
A credential is present in a process, or it is not
Strict runtime identity split across services & jobs
Plugins and file parsers execute in confined processes with zero warehouse credentials. Each store operation belongs to one narrow runtime identity.
Irreversible actions require two independent people
Transaction-level separation of duties
Data erasure and key destruction require a second approver from an independent grant lineage. Managing users and approving erasures are mutually exclusive permissions.
Built-In AI, Universal MCP & Automations
How Keystone powers natural-speech Q&A, any LLM via MCP, and AI-built dashboards & automations
Giving an external LLM a raw database connection string leaks restricted rows and produces made-up numbers. Project Keystone gives your built-in LLM and your external MCP-connected models (OpenAI, Anthropic Claude, Google Gemini, or locked-down in-house LLMs) a governed execution layer:
- 1. Universal MCP Server & Built-In Natural Speech AI: Connect Claude, OpenAI, Gemini, or self-hosted models via MCP, or use the built-in LLM in natural speech. Every call runs as the caller’s own RBAC Principal.
- 2. AI Builder for Dashboards, Reports & Automations: Ask the AI in plain language to construct live interactive dashboards, schedule custom reports, or create automated workflows and threshold triggers.
- 3. Verified Spec & Figure Check: The semantic engine computes the math; every number, unit, and date in an AI response or report is verified against the semantic layer before execution.
# 1. LLM & Principal Identity Check
client: mcp/claude-enterprise | builtin-llm | local-llm
principal: role/exco_operator · pii_reveal: masked
# 2. Natural Speech Intent → Governed Action
action: build_dashboard + create_automation_trigger
sources: [polymarket, kalshi, news_wire, erp_orders]
accounting: read(1,420,880) = ins(1,420,812) + skip(68) + exc(0)
# 3. Output & Automation Verification Gate
figures_checked: 100% match semantic spec
audit_logged: dashboard_built · trigger_armed · audit #98412
Zero-Trust Runtime Confinement
Least-privilege identity matrix across every store operation
A component is not a security boundary if everything shares one database connection string. Project Keystone splits runtime identities so each process holds only the narrowest grant for its single job.
| Runtime Identity | Responsibility | Permitted Reach | Strict Confinement Boundary |
|---|---|---|---|
| Extraction Workers | Short-lived source readers & file/stream parsers | Assigned external source + write-only raw landing prefix | Zero analytical store or operational control-plane credentials. Untrusted parser code runs confined. |
| Registration & Accounting Job | Validates landed batches & enforces run accounting | Appends to Bronze schema, Exception Ledger, and control.run | Cannot modify Silver/Gold models, alter schemas, or delete records. |
| Model Builder Job | Builds Silver & Gold models + Identity Spine | Reads Bronze & writes rebuildable Silver/Gold schemas | Cannot bypass retention rules or alter user permission grants. |
| Governed Query & MCP Service | Executes semantic queries for built-in AI, dashboards & external LLMs via MCP | Reads Gold views strictly under per-principal RBAC row & column policies | Cannot write, mutate, or bypass row-level security. |
| AI Builder & Automation Engine | Natural speech Q&A, dashboard/report builder & workflow automations | Calls Governed Query Service via semantic metric specs and executes scoped automations | Database credentials never enter LLM context; any unverified number rejects the output. |
| Retention & Erasure Job | The single identity permitted to delete expired or erased rows | Executes scheduled expiry & two-person approved POPIA/GDPR erasures | Every deletion is logged in the immutable deletion ledger. |
Sovereign Deployment Profiles
Same platform across On-Premise, Cloud, and Hybrid
On-Premise / Private Infrastructure
Maximum data sovereignty inside your own walls—with in-house locked-down LLMs.
Deploy Project Keystone directly onto your own Linux servers, VMware/Proxmox cluster, or air-gapped private data center. Pair it with your own locked-down local LLMs so not a single byte of data or AI prompt ever leaves your building.
- Zero data leaves your private network boundary
- Connects to in-house, locked-down local LLMs via built-in UI or MCP
- Single-server Docker profile or multi-node orchestrated cluster
- Full hardware and cryptographic key control
Dedicated Cloud Deployment (AWS / GCP / Azure)
Stood up in your cloud account, not a shared vendor SaaS.
We provision Project Keystone inside an AWS, Google Cloud, or Azure account that your organization owns—using your KMS encryption keys, your VPC/networking rules, and your chosen cloud or enterprise LLMs (OpenAI, Claude, Gemini).
- Provisioned via automated infrastructure-as-code in your account
- Encrypted at rest with your own customer-managed KMS keys
- Multi-zone high availability for queries, dashboards, and automations
- Optional full handover: revoke our deployment role when ready
Hybrid & Vexalabs-Managed
Keep sensitive stores local while scaling compute & AI in the cloud.
Keep raw PII, subject-reference maps, or legacy database extractors inside your private environment while running analytical models, market/news ingestion, or cloud LLMs in a dedicated enclave—operated by your team, Vexalabs, or together.
- Strict egress policy & encrypted service-to-service tunnels
- PII pseudonymisation before data crosses environment boundaries
- Co-managed with Vexalabs Outsourced CTO & Data Engineering support
- Seamless migration path between on-prem and cloud profiles
Want to review the Project Keystone architecture with our team?
Get in touch with our engineering leadership to discuss deploying Project Keystone in your environment, cloud, or hybrid setup.