CloudSense AI

Semantic Execution Layer for Enterprise AI.

Enterprise AI that doesn’t hallucinate. One typed semantic graph. Governed SQL at compile time.

CloudSense encodes your business meaning — metric definitions, entity relationships, governance scopes — and emits deterministic, auditable SQL. No hallucination. No unauthorized queries reaching the warehouse.

31%
Raw NL2SQL
87%
With metric layer
98–100%
With CloudSense

The cloud backbone every edge decision depends on.

The problem

Every enterprise has this. No one planned it.

5 disconnected implementations · 3 incompatible auth models · 0 shared governance · Every new use case starts from scratch.

NL2SQL POC sprawl

Three teams built NL2SQL on the same warehouse. Different revenue definitions. Different auth. None share governance. A fourth team is starting another one.

The MCP Server trap

"Just register more MCP servers" is the advice. But who joins results across servers? Not the LLM — token cost is unbounded, joins can't be audited, and your raw data crosses a third-party inference API.

Advisory governance fails

Unauthorized queries reach the warehouse — filtered after the fact. AI agents bypass rules by rephrasing. Governance failures show up in logs after the damage is done.

31% NL2SQL accuracy

Raw NL2SQL on Spider 2.0 benchmarks at 31%. Without a semantic layer encoding business meaning, models invent metric definitions, join wrong tables, and return confidently wrong results.

These aren’t model problems. They’re semantic execution problems.

Vendor clarity

Everyone claims a semantic layer. Most don’t have one.

Your data warehouse added metadata tags. Your catalog added AI search. Neither executes governed queries.

Data WarehouseData CatalogCloudSense
JobStore & process data at scaleIndex & document data assets Encode meaning + emit governed SQL
Query executionYes — direct SQLNo — documentation only Emits governed SQL to warehouse
GovernanceAdvisory / post-executionMetadata tags & lineage Compile-time RBAC/ABAC
Semantic layer?❌ Metadata tags ≠ governed SQL❌ Lineage graphs ≠ governed SQL ✅ Single definition per metric
AI-ready?⚠ Adds NL2SQL on top — still 31%⚠ Context but no execution ✅ 98–100% on covered questions

How it works

One semantic graph. Every consumer. Zero hallucination.

CONSUMERSAI Agents · BI Dashboards · MCP Tools · EdgeSense · REST Clientsgoverned SQL · compile-time policiesCLOUDSENSESemantic GraphPostgreSQL + AGEopenCypher · pgvectorFederated QueryPG FDW · 16+ dialectsHeterogeneous sourcesCompile-Time RBACABAC injected beforewarehouse executionMCP + RESTAgent-native · stdioSSE · HTTP · JDBCMAP → DEFINE → RESOLVE → COMPILEautonomous drift detection · typed semantic graph · one definition per metricJDBC / ODBC / MCP / REST · raw dataDATA SOURCESSharePoint · AWS S3 · PostgreSQL · ServiceNow · Salesforce · MCP Sources
1MAP

Introspect all connected sources — schemas, columns, relationships

2DEFINE

Encode business meaning: metric formulas, entity definitions, governance scopes

3RESOLVE

Ground every query term in the semantic graph — no invented definitions

4COMPILE

Emit native SQL with RBAC/ABAC injected — deterministic, auditable

Why it works

Compile-time governance. Not advisory. Enforced.

⚠ Advisory governance — every POC today

Unauthorized queries reach the warehouse — filtered after the fact
AI agents bypass rules by rephrasing queries
Governance failures visible only in logs after the fact
Manual catalog curation — definitions drift silently

✅ Compile-time governance — CloudSense

RBAC + ABAC predicates injected during SQL generation
Unauthorized queries fail to compile — never hit the warehouse
Every query: reproducible audit trail + proof of access
Autonomous drift detection keeps definitions current

EdgeSense AI

The semantic layer. To the last mile.

CloudSense governs the cloud. EdgeSense extends the semantic graph to field devices, factories, and air-gapped environments.

How semantics travel to the edge

CloudSense

Full semantic graph

All enterprise entities, metric definitions, governance scopes, and relationships

scoped slice

signed + encrypted

EdgeSense

Local semantic graph

Role-scoped slice in sqlite — only the entities and rules this device and role need

escalate

if needed

On-device SLM

Grounded inference

Model resolves against local graph first — escalates to CloudSense only when confidence demands it

CRDT sync keeps the local graph current when connectivity returns. Outcomes feed back to CloudSense to strengthen the shared semantic graph.

📡

No connectivity

Field tech offline — NL2SQL POC returns 503. Technician uses a PDF manual. 45 minutes of search.

EdgeSense answers from local sqlite semantic graph in < 2s

🏭

Latency

Cloud round-trip: 3–8s per query. Line runs faster than AI responds. $2.4M system, unused.

SLM inference local — sub-second on covered questions

🔒

Data sovereignty

SCADA telemetry is CIP-classified. Cloud LLM cannot see it. Every query sanitised by hand.

CIP-tagged data stays on device. Only non-classified content reaches cloud.

EdgeSense stack

Gemma 3n · Qwen2.5-VL SLMs
sqlite + SQLCipher AES-256
CRDT sync · MCP Client → CloudSense
NERC CIP: classified data never leaves device

Stop building POCs. Start shipping governed AI.

Thinking Sense · CloudSense AI · Early access 2026

Get early access →