Who it is for
Teams that need governed data questions
Organizations that want natural-language-to-SQL, semantic typing, and provenance checks inside a controlled deployment boundary rather than a public AI database agent.
ENGAGEMENT-LED · SEMANTIC SQL
Zerde turns governed data questions into reviewable SQL.
Zerde is semantic SQL infrastructure for customer-owned environments. It converts plain questions into reviewable SQL candidates against known tables, uses schema and provenance context to constrain the work, and is scoped per engagement because model choice, schema design, review rules, and failure tolerance are customer-specific.
What it does
Constrains and reviews semantic SQL
Turns plain-language questions into reviewable SQL candidates against known tables, applies type and provenance context, and keeps model/runtime choices inside the scoped environment.
What it does not do
Not an unmanaged chatbot for production data
Zerde is not a public chatbot, open-ended autonomous database operator, generic BI replacement, or license to execute every generated query without review rules.
Qualify the engagement with these details
Schema and sources
Tables, columns, joins, row-level rules, data sources, sensitive fields, and how schema changes should be handled.
Query classes
Allowed question types, disallowed operations, write permissions, review thresholds, and failure behavior.
Model and runtime
Local model preference, llama.cpp backend, GPU/CPU target, memory constraints, service boundary, and administrator access.
Governance tolerance
False-positive tolerance, audit requirements, provenance requirements, human approval flow, and support window.
Natural-language to SQL
Grammar-constrained inference narrows model output to reviewable SQL candidates that can be validated, rejected, retried, or reviewed before use.
Five-tier type cascade
Semantic type inference resolves columns through explicit, structural, statistical, contextual, and LLM tiers — escalating only when needed.
BLAKE3 provenance chain
Transformation records can be chained and verified so tampering or mismatch breaks visibly instead of silently.
GPU detection with CPU fallback
CUDA and Vulkan backends can be scoped where available; CPU fallback is part of the deployment conversation when no accelerator is present.
RAG-assisted retry path
Failed SQL execution can retrieve related context and attempt a bounded retry when the engagement allows that behavior.
Local service, embedded UI
The deployment can expose a local HTTP service and browser UI without turning the data boundary into a hosted SaaS surface.
What happens after contact
- Fit check: confirm whether Zerde belongs in the workflow and where human review must sit.
- Written scope: define schemas, allowed query classes, model/runtime choices, endpoint exposure, review rules, and support terms.
- Delivery: provide the scoped service, configuration notes, deployment instructions, and support channel agreed in scope.